The sandboxed Pyodide host (used when ENABLE_PYODIDE_FILE_PERSISTENCE is
disabled, the default) embeds its script in a String.raw template. The
matplotlib show() override was written with '\\t' escapes as if in a normal
string context, but String.raw preserves them verbatim, so the iframe's
script parser turns them into literal backslash-t characters in the
generated Python source. Pyodide then fails to compile any code that
triggers the matplotlib patch with "SyntaxError: unexpected character
after line continuation character", which is why matplotlib only worked
with the file persistence worker path enabled.
Use single '\t' escapes instead: String.raw keeps them as-is in the
script text and the sandbox's JS parser produces real tab indentation,
matching the working implementation in pyodide.worker.ts.
Fixes#26660
Assistant responses are now stored as structured output items on message.output, with message.content left empty. The action payload built in chatActionHandler still sent message.content only, so action functions received assistant messages with an empty content property. Derive the content from the structured output via getOutputText, falling back to message.content, matching how the rest of Chat.svelte resolves assistant text.
Fixes#26672
_ssrf_safe_new_conn unpacks each entry of self.socket_options straight into socket.setsockopt(), which accepts exactly 3 positional arguments. urllib3-future, a drop-in fork that shadows the urllib3 package whenever it is installed (for example as a dependency of niquests pulled in through a tool or function's requirements), declares its default socket options with a per-protocol 4th element: [(socket.IPPROTO_TCP, socket.TCP_NODELAY, 1, "tcp")]. Its own _set_socket_options() strips that element before calling setsockopt(), but our override does not, so with urllib3-future present every synchronous web fetch (fetch_url, web search loading) fails on connect with "TypeError: setsockopt() takes exactly 3 arguments (4 given)" and returns empty content.
Mirror urllib3-future's handling in the override: for 4-element options whose last element is a protocol string, apply "tcp" options truncated to the first 3 elements and skip "udp" options (all sockets created here are SOCK_STREAM). Plain 3-element options, and any other shapes stock urllib3 would accept, are passed through unchanged, so behavior with stock urllib3 (which only ever uses 3-element tuples) is identical.
Verified locally: with urllib3-future installed the loader previously raised the TypeError on every URL and now fetches successfully; with stock urllib3 2.3.0 and 2.7.0 fetches behave the same before and after.
Note: #26015 reported this same crash but attributed it to stock urllib3 2.x, which only uses 3-element tuples; the 4-element form comes from urllib3-future shadowing urllib3.
Fixes#26791
is_regex_pattern only recognized the BRE-escaped form \| and not a bare |,
so a pattern like "Jornak|Silverlake|Orissa" was treated as one literal
string (including the pipe characters) and silently returned no matches.
This contradicted the tool docstring, which explicitly advertises
"error|warn" as an auto-detected regex example, and misled models into
concluding the searched terms were absent from the file.
Checking for a bare | also covers the escaped form, since \| contains |,
and normalize_regex already converts escaped pipes before compilation.
Literal patterns without regex metacharacters are unaffected.
Fixes#26781
SafeWebBaseLoader._unpack_fetch_results assigned the resolved parser to the parser parameter itself, so the None check only ran for the first URL. In a mixed batch every later document was parsed with whatever the first URL happened to select: an .xml feed first meant all following HTML pages went through the xml parser (broken text extraction), and an HTML page first meant .xml URLs were parsed as HTML. Web search regularly fetches mixed batches, so this silently degraded extraction quality depending on result order.
The parser is now resolved per URL; an explicitly passed parser still applies to the whole batch as before. Verified with mixed xml/html batches in both orders and with an explicit parser override.
When WEBSOCKET_MANAGER=redis, app.state.MODELS and the socket session/
usage pools are Redis-backed dicts, so every membership test and
getitem is a network round trip:
- generate_chat_completion checked `model_id not in models` (HEXISTS)
and then read `models[model_id]` (HGET) on every chat completion.
A single .get() now serves both, with the same not-found error.
- The direct-connection branch spread the pool with `{**MODELS, ...}`,
which iterates keys() then fetches each value — HKEYS plus one HGET
per model. dict(MODELS.items()) issues a single HGETALL instead.
- get_user_ids_from_room called SESSION_POOL.get(sid) twice per
session (once to filter, once for the value); the usage handler
checked membership then fetched the same key. Both now do one
lookup.
In non-Redis mode these are plain dicts and behavior is identical.
Claude-Session: https://claude.ai/code/session_01MHg5zs1VBjvRWQ54qHpfYD
Co-authored-by: Claude <noreply@anthropic.com>
The streaming handler's content accumulator is a closure cell (declared nonlocal in stream_body_handler), and CPython's in-place string append optimization only applies to plain local variables (STORE_FAST), never to cell variables (STORE_DEREF). The content += value form introduced in #27231 therefore still allocates and copies the full accumulated string on every delta, exactly like the f-string it replaced; whether that copy is cheap or expensive is up to the allocator, and measurements swing accordingly (6 to 83 ms of pure copying for a 400 KB response on Python 3.12, against 0.3 ms for this patch).
Accumulate the deltas in a list instead and join once at the single read site (publish_chat_finished_event at stream end). List append is amortized O(1) with no dependence on reference counts, bytecode specialization or allocator behaviour, so accumulation is O(n) by construction. The non-str fallback keeps the previous f-string coercion semantics.
Verified end to end against a mock SSE upstream: a streamed chat with a think-tag block plus 40 content deltas produces output items, message text, reasoning text and usage identical to current dev, with no errors in the server log.
CompressMiddleware was registered before AuditLoggingMiddleware. Starlette prepends on add_middleware, so the audit layer ended up outside compression and, at the REQUEST_RESPONSE level, recorded the zstd/brotli/gzip bytes of every response, decoded with errors='replace'. Any client that sent Accept-Encoding (i.e. every browser) therefore produced audit entries whose response_object was unreadable mojibake.
Registering the audit middleware before the compression middleware places it inside compression, so it observes the response body exactly as the route produced it while the client still receives the compressed stream.
Verified with a stacked ASGI harness: in the old order the captured body is not parseable; in the new order the captured body round-trips as the original JSON and the client response stays compressed.
The header value was truncated with int(), so every request faster than one second reported X-Process-Time: 0 and the header carried no information for exactly the requests it is meant to describe. Emit fractional seconds with microsecond precision instead, matching the pre-ASGI-refactor behavior where the raw float was sent.
With audit logging enabled, every audited request authenticated twice. The route dependency resolved the user once, and then _log_audit_entry called get_current_user again in the request's finally block: a second JWT decode, two more Redis revocation lookups, a second user row fetch with pydantic validation and, crucially, a second fire-and-forget last-active write transaction per request.
get_current_user now stashes the resolved user on the scope-backed request state (the same mechanism the auth middleware already uses for request.state.token), and the audit middleware reuses it, falling back to the old resolution only when no user was stashed (e.g. routes without an auth dependency). While in the file, the audit path patterns are compiled once in the constructor instead of per request, and the always-log endpoint set is a class attribute instead of a per-call literal; both are fixed for the process lifetime.
Benchmark:
| metric | before | after |
| --- | --- | --- |
| audit auth resolution, CPU floor (JWT decode + user validate only) | 16.7 us | 0.24 us |
| extra work per audited request | 2 Redis GETs + 1 user SELECT + 1 last-active write | none |
The before column understates the saving: it excludes the Redis and DB round trips listed in the second row, which dominate in real deployments.
Functionally verified with a stacked ASGI harness: when the route resolves a user the audit entry carries that user and the auth pipeline is not invoked again; without a stashed user the fallback path still resolves and logs correctly; the skip matrix (exclusions, whitelist mode, always-log auth endpoints, unauthenticated and non-audited methods) is unchanged.
Both session factories run with expire_on_commit=False, so ORM objects keep their attribute values after commit. Every session.refresh issued right after a commit therefore re-SELECTed a row whose values the session already held, including full chat JSON blobs and user settings, purely to overwrite identical data. Fifty such calls existed across the model layer, covering nearly every write path in the app (chat inserts, title updates, pin/archive toggles, user role and settings updates, tool, prompt, function, model, file, tag, feedback, memory, automation and grant writes).
All fifty are removed. The only refreshes with an actual job were the two update-then-reload paths in tools and skills, where a Core UPDATE statement bypasses the identity map; those now use session.get(..., populate_existing=True), which guarantees a fresh row in one SELECT whether or not the row was already present in the session (the previous code issued get plus refresh, two SELECTs, on the default configuration).
Benchmark (real SQLite DB, per write):
| write path | before | after |
| --- | --- | --- |
| chat title update, ~600 KB chat blob | 2.08 ms | 1.24 ms |
| user role update, small row | 1.21 ms | 0.68 ms |
On Postgres each removed refresh is additionally a network round trip. The chat-blob case also skips re-parsing the entire JSON document per write.
Functionally verified against a fresh database: user insert, role and settings updates, chat insert (including the server-default meta column, which is always provided client-side), title update and pin toggle, tool insert and the Core-update reload path, tag insert and the prompt insert flow that pins version_id after history creation all return correct values and persist correctly.
get_all_models runs on every models refresh and, without the base-models cache (off by default), on every /api/models request. Several of its costs multiplied by the model count for no reason:
- The active action and filter id sets were derived from get_functions_by_type, which loads full function rows including plugin source and validates them, only for the ids and is_global flags. A generalized column-only query now returns (id, is_global) tuples; the existing filter-specific helper delegates to it.
- Action priorities were computed inside the per-model sort key, constructing a pydantic Valves object per action per model; with global actions in every model's list that was models x actions constructions per refresh. Priorities are now memoized per action.
- Global action and filter item dicts were rebuilt per model from the same modules. The item lists are now built once per function and shallow-copied per model, keeping per-model dicts independent exactly as before (nested values were already shared).
- Deactivated base-model overrides were dropped with models.remove, a linear scan and shift per removal; removals are now collected and filtered out in one identity-based pass, preserving list.remove's exact object semantics.
- RedisDict.set fingerprinted the payload by serializing the already-serialized mapping a second time plus a sha256; a direct dict comparison against the last written mapping has the same skip semantics without re-serializing anything.
- /api/models did tag normalization and profile-image stripping for every model before access filtering discarded the invisible ones, and always evaluated a json.dumps debug f-string; the work now runs only on visible models and the debug line is gated on the log level. The duplicate-id dedup keeps its position before filtering so the effective-model semantics are unchanged.
Benchmark:
| metric | before | after |
| --- | --- | --- |
| model-cache fingerprint, 200 models | 45 us | 1.4 us |
| action priority Valves builds, 200 models x 4 global actions | 0.37 ms (800 builds) | 0.002 ms (4 builds) |
| function-table payload for id sets | full rows incl. source | (id, is_global) tuples |
Functionally verified: the column-only id query matches the full-row query for actions and filters including inactive exclusion, and the fingerprint skip logic writes on first set, skips identical payloads, updates plus deletes stale keys on change and clears on empty, against a scripted fake Redis.
Tools.get_tools(defer_content=True) contained the literal dead statement "stmt = stmt": the deferral was a no-op, so every tools listing loaded the full Python source of every tool (five caller sites pass defer_content=True expecting the optimization: the tools list endpoints and the user and group permission overviews). The listing now selects every column except content, and ToolModel.content becomes optional to represent deferred rows; router projections are content-less response models, so nothing downstream reads the source on these paths.
On top of that, get_tools_by_user_id issued one grant query per non-owned tool and Knowledges.get_knowledge_bases_by_user_id did the same per knowledge base (the latter also sits inside per-file access checks). Both now resolve grants for all non-owned rows in a single get_accessible_resource_ids call, the same batch helper the model listing already uses.
Benchmark (real SQLite DB):
| metric | before | after |
| --- | --- | --- |
| tools listing, 33 tools x ~200 KB source | 5.13 ms | 3.18 ms |
| grant queries per accessible-tools call, N non-owned tools | N | 1 |
| grant queries per accessible-KBs call, N non-owned KBs | N | 1 |
The listing row scales with source size; on Postgres the deferral additionally avoids shipping every tool's source over the wire per listing, and each removed grant query was a real round trip.
Functionally verified: deferred listings match full listings field for field with content None, grants included and router projections working; access filtering returns exactly owned plus granted tools and knowledge bases and nothing for strangers; full (non-deferred) reads still carry the source.
Knowledges.get_file_metadatas_by_id fetched full File rows, whose data column carries the entire extracted text of each document, validated each into a FileModel and then threw everything except id, hash, meta and the timestamps away. The function backs every knowledge base detail view and runs again after every file add or remove (eight call sites in the knowledge router), so rendering a filename list for a 50-file knowledge base parsed tens of megabytes of JSON per request.
The listing now selects exactly the five columns the response needs, joined through KnowledgeFile, mirroring the column-only helper that already existed in the files model for id-based lookups.
Benchmark (real SQLite DB, 50 files with ~200 KB extracted text each):
| metric | before | after |
| --- | --- | --- |
| KB file metadata listing | 14.2 ms | 1.20 ms |
The gap widens linearly with file size and count since extracted contents no longer get read, parsed or validated at all.
Functionally verified: output matches the old implementation field for field on all 50 files and an unknown knowledge base still returns an empty list.
has_access_to_file runs for every non-owner file GET, per RAG file check and per shared-chat or model-attached file. Its final step called Models.get_models_by_user_id, which issued one grant query per non-owned workspace model, so a single file check on an instance with M workspace models cost M grant queries plus a group query, with the deny path always paying full price. Its collection_name step listed every knowledge base the user can access (itself one grant query per knowledge base) just to scan the list for one id. And get_accessible_folder_files repeated the whole pipeline per folder entry, refetching the caller's group memberships every time.
Three changes, all using parameters and helpers that already exist:
- Models.get_models_by_user_id resolves grants for all non-owned models in one get_accessible_resource_ids call and accepts prefetched user_group_ids.
- The collection_name check fetches the one referenced knowledge base and performs a single owner-or-grant check with the already-resolved group ids, preserving the write-requires-owner guard exactly (including its short-circuit before any grant query).
- get_accessible_folder_files resolves group ids once and threads them through every per-entry check.
Benchmark:
| metric | before | after |
| --- | --- | --- |
| filter loop CPU, 300 workspace models (queries stubbed) | 47 us | 19 us |
| grant queries per file-access check, M workspace models | M | 1 |
| group membership queries per folder listing, F files | F | 1 |
The stubbed CPU row understates the win: each removed query in the other two rows was a real database round trip.
Functionally verified with stubbed accessors: owned plus granted models are returned with owned ids excluded from the batch query; model-attached file access resolves through the batched path; the collection_name path does one KB fetch and one grant check with no full listing; a missing KB falls through; write access via a KB still requires the KB owner to own the file and short-circuits before the grant query; folder listings fetch groups exactly once.
Two independent sources of fixed per-request cost:
The async SQLite engine was created with pool_pre_ping=True. A pre-ping guards against server connections dropped by timeouts or restarts, which cannot happen to a local SQLite file; each ping still costs a hop into the aiosqlite worker thread plus a SELECT 1 on every connection checkout, and with session sharing off a single request checks out a connection for every model-layer call it makes. The Postgres engines keep their pre-ping, where it is actually protective.
CommitSessionMiddleware unconditionally ran ScopedSession.commit() plus remove() after every HTTP request. The scoped registry instantiates a session on first access, so on the vast majority of requests (which never touch the sync session, per the middleware's own docstring) this built a Session, opened and committed an empty transaction and tore everything down for nothing. The middleware now checks ScopedSession.registry.has() first: requests that used the sync session are committed and removed exactly as before, on success and on the rollback path alike, and idle requests skip the machinery entirely.
Benchmark (real SQLite database):
| metric | before | after |
| --- | --- | --- |
| user row fetch incl. session + connection checkout | 681 us | 514 us |
| idle-request sync session work (create + empty commit + teardown) | 12.3 us | 0.26 us |
The first row saves per model-layer call, not per request: a request making five DB calls saves the checkout ping five times.
Functionally verified: normal reads and writes work with pre-ping off; an idle request through the middleware leaves no sync session behind; a request that uses the sync session still gets committed and removed.
The branch that normalizes plain JSON error lines from streaming upstreams (lines without the SSE data: prefix) called Chats.upsert_message_to_chat_by_id_and_message_id without await. The coroutine was never executed, so the error was never written to the chat and Python emitted a "coroutine was never awaited" RuntimeWarning instead. The frontend still received the error event, but after a reload the message showed no trace of the failure.
The parallel error-persist branch further down the same handler already awaits the call; this aligns the two.
Several spots in the chat pipeline issued sequential single-key config
SELECTs, or fetched the same key twice back-to-back, on every request:
- chat_completion_tools_handler: task model default/external and the
tools prompt template were four sequential Config round trips (the
template was fetched twice). One batched Config.get_many now serves
all of them.
- chat_completion_files_handler: the six RAG settings (top_k,
top_k_reranker, relevance_threshold, hybrid_bm25_weight,
enable_hybrid_search, full_context) were six sequential round trips
inside the retrieval call. Batched into one get_many.
- Voice and code-interpreter prompt templates were each fetched twice
within one conditional; fetch once and reuse. The code-interpreter
engine was likewise fetched twice per execution.
- Skill resolution fetched the accessible-skills list, kept only the
ids, then re-fetched each mentioned skill by id (N+1). Reuse the
rows from the access query.
Value semantics are identical: get_many applies the same defaults as
the individual gets, and the pre-existing truthiness/empty-string
checks on templates are preserved exactly.
Claude-Session: https://claude.ai/code/session_01MHg5zs1VBjvRWQ54qHpfYD
Co-authored-by: Claude <noreply@anthropic.com>
The middleware code-interpreter path defines `restricted_import` as `async def` and assigns it to `builtins.__import__`, which Python's import machinery calls synchronously. Calling an async function returns a coroutine without running its body, so the blocklist check never executes and `_real_import` is never called. When `CODE_INTERPRETER_BLOCKED_MODULES` is set, blocked modules are therefore not blocked, and every subsequent import inside the interpreter binds a dangling coroutine instead of the module, breaking legitimate imports as well.
Define the hook as a regular `def`, matching the working implementation in `tools/builtin.py`. A blocked top-level import now raises `ImportError`, and all other imports pass through to the real importer.
The per-request Ollama handlers (chat, generate, embed, embeddings,
and the OpenAI-compat completions/chat-completions/messages/responses
endpoints) fetched 'ollama.api_configs' up to three times and
'ollama.base_urls' separately within a single request — the .get()
default-argument pattern made the second api_configs fetch
unconditional, and get_api_key() triggered a third. Up to four
sequential SELECTs per request collapse to one.
A new get_ollama_connection_config() helper fetches base_urls and
api_configs together in one batched Config.get_many where both are
needed; handlers that only need api_configs fetch it once into a
local. Admin operations (pull/push/copy/delete) and the TTL-cached
model-list path are deliberately left untouched.
Resolution semantics (str(idx) key first, url-key legacy fallback,
same defaults) are unchanged.
Claude-Session: https://claude.ai/code/session_01MHg5zs1VBjvRWQ54qHpfYD
Co-authored-by: Claude <noreply@anthropic.com>
SecurityHeadersMiddleware called set_security_headers() on every
response — 14 os.environ.get lookups plus a regex validation per
configured header, for values that are static for the process
lifetime. Compute the header list once at construction; when no
security env vars are set, skip wrapping send entirely.
RedirectMiddleware decoded and parse_qs'd the query string of every
GET, though it only acts on /watch?v= and ?shared= URLs. Add a cheap
path/substring precheck first; a false positive just falls through to
the previous full parse, so no redirect behavior changes.
Verified byte-identical responses (status, Location, header values)
against the previous implementations across redirect, passthrough,
and no-env cases.
Claude-Session: https://claude.ai/code/session_01MHg5zs1VBjvRWQ54qHpfYD
Co-authored-by: Claude <noreply@anthropic.com>
The streaming handler rebuilt the full accumulated response with
`content = f'{content}{value}'` on every content delta — a complete
string copy per chunk, making accumulation O(n^2) over the response
length. Use in-place `content += value` for the (universal) str case,
which CPython extends in place, keeping accumulation O(n); the
f-string fallback is preserved for non-str values so coercion
behavior is unchanged.
In the ENABLE_REALTIME_CHAT_SAVE branch, full_output() — which
concatenates the entire accumulated output — was called twice per
chunk (once for the DB upsert, once for the emitted delta). Compute
it once and reuse.
Claude-Session: https://claude.ai/code/session_01MHg5zs1VBjvRWQ54qHpfYD
Co-authored-by: Claude <noreply@anthropic.com>
The Automatic1111 branch of the image generation route called set_image_model whenever the request carried a model field. On this backend set_image_model is not request-scoped: it persists image_generation.model to the global configuration and posts the new sd_model_checkpoint to the shared server, because Automatic1111 holds a single checkpoint instance-wide. A non-admin holding features.image_generation could therefore change the instance-wide image model and the shared backend checkpoint for every user by sending a model on an ordinary generation request, even though the setting is otherwise managed only through the admin-only image configuration route and the frontend never sends this field.
Gate the switch on an admin caller. A non-admin now generates on the currently configured checkpoint and the model field no longer mutates global state; admins retain per-request model switching here and through the image configuration route. Image editing is unaffected, as it selects its model per request without writing global configuration.
The chat action route loaded a Function by its raw action_id and executed its action callable after only checking that the id and the requested model existed. The model list that the client renders actions from resolves each model's actions to the active action-type Functions that are global or assigned to that model, and the action route did not mirror that resolution, so a disabled, unassigned, or wrong-type Function, or an action on a model the caller cannot access, could be reached by calling the route directly.
Gate the route on the same rules the model resolution applies: the Function must be an active action, and for server-resolved models the caller must have model access and the action must be one the model actually surfaces (matched by function id, the prefix of each model actions entry, so single and sub-actions both resolve). Direct connections carry a client-supplied model the caller already owns, so the model-bound checks are scoped to non-direct calls; the active-action check always applies. Executing admin-authored Function code remains intended behaviour — this only keeps the route consistent with which actions each model exposes.
Co-authored-by: komyunghan <komyunghan@users.noreply.github.com>
get_all_models ran four function-table queries: global actions, active
actions, global filters, active filters. Global functions are by
definition (type, is_active=True, is_global=True) — a subset of the
active set — so the global id sets can be derived from the active-rows
queries' is_global flag. Four queries become two, and each dropped
query returned full rows including every plugin's source code.
Also folds the mid-function 'models.default_metadata' read into the
Config.get_many already issued at the top of the function (one fewer
round trip; the existing `or {}` default handling is preserved).
Claude-Session: https://claude.ai/code/session_01MHg5zs1VBjvRWQ54qHpfYD
Co-authored-by: Claude <noreply@anthropic.com>
* perf: cut per-instance CPU cost of shared socket.io Redis pub/sub channel
Profiling a multi-instance deployment (py-spy --gil) showed ~44% of worker
CPU in the socket.io pub/sub listener. Two causes, two fixes:
- Add hiredis so redis-py parses the RESP protocol in C instead of pure
Python (redis/_parsers/resp3.py alone accounted for ~28% of GIL samples;
redis-py auto-selects the hiredis parser when importable).
- Subclass AsyncRedisManager to drop emits whose target room has no local
participants before upstream _handle_emit re-encodes the full packet.
Every instance receives every emit published on the shared channel, so
with N instances all but the hosting one were paying full packet
re-serialization per message just to deliver it to nobody. Broadcasts
(room=None) are unaffected.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Q9CQ9qnp3sZGYQQwztsCJT
* fix: restrict pub/sub emit early-out to string rooms
Adversarial review against python-socketio 5.16.2 found one divergence
from upstream: for a degenerate empty-sequence room (emit to room=[]) on
an instance whose namespace has no local clients, the filter's
get_participants probe raises IndexError from room[0] where upstream
returns silently at the namespace guard and still publishes to Redis.
Open WebUI only ever emits to scalar string rooms or room=None, so the
case is unreachable today; guard on isinstance(room, str) anyway so any
non-string room shape passes through to upstream behavior unchanged.
Every open-webui emit uses a string room, so the fast path still covers
all real traffic.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Q9CQ9qnp3sZGYQQwztsCJT
* Update requirements.txt
* Update pyproject.toml
* Update requirements-min.txt
---------
Co-authored-by: Claude <noreply@anthropic.com>
When an upstream provider rejects a request (e.g. a 400 for a
max_tokens value above the model's ceiling), the actionable error
message was only published to event sinks, which are invisible unless
an event function or webhook is configured. Admins had to query the
provider's API directly to diagnose failures (open-webui#27237).
Add a single log line in publish_model_provider_request_failed — the
chokepoint every upstream failure path (OpenAI-compatible chat,
embeddings, responses, token counting, and Ollama) already routes
through — recording status, provider, url, model, error code, and the
upstream message truncated to 1000 chars. 4xx logs at WARNING, 5xx at
ERROR. Client-facing responses are unchanged, so no additional error
detail is exposed in the chat.
Claude-Session: https://claude.ai/code/session_018VecyiPejru1EVF5yfe2sU
Co-authored-by: Claude <noreply@anthropic.com>
`sharing.folders` is present in `DEFAULT_USER_PERMISSIONS` but absent from the `SharingPermissions` response/update schema, so the admin default and group permission API silently drops it on every round-trip and the setting is never saved. Add `folders: bool = False`, matching the config default (`USER_PERMISSIONS_FOLDERS_ALLOW_SHARING`), restoring parity with `DEFAULT_USER_PERMISSIONS`.
Fixes#27120.
generate_function_chat_completion rewrote form_data['model'] to the base
model id in place. Because that dict is the same object process_chat_response
later re-submits for the post-tool-call continuation, the continuation was
access-checked against the base model instead of the user-facing preset.
For a public preset over a private/unregistered pipe base, a non-admin's
first message succeeded but the continuation after a native tool call was
denied and silently swallowed, aborting the response (admins skip the
user-only check, so they were unaffected).
Operate on a shallow copy of form_data, mirroring the openai/ollama routers
which already substitute the base model on a copy. The continuation now
re-checks the preset the user actually has access to.
Claude-Session: https://claude.ai/code/session_018toPfJW1hMXAhokGaL43Ep
Co-authored-by: Claude <noreply@anthropic.com>
When a chat is started from the home page, the URL is switched to /c/{id} with
history.replaceState, so the Chat component is never remounted and chatIdProp
stays empty for the lifetime of that view. Both websocket recovery paths,
handleSocketConnect and the chat:active fallback, are gated on chatIdProp and
therefore never run for these chats. After any websocket drop during a response
(mobile backgrounding, VPN or IP change, wake from sleep) the completed response
is never fetched and the chat stays stuck in a loading state until a manual page
refresh. Chats opened directly via /c/{id} recover fine, which is why the bug
only reproduces reliably on freshly started chats.
Gate both recovery paths on the chatId store instead, which is set for every
persisted chat, and let loadChat fall back to it so the recovery reload also
works when chatIdProp is empty. Chats opened via /c/{id} behave exactly as
before and temporary chats stay excluded.
The same gate likely explains the remaining reports in #26315.
Fixes#26844
The system-terminal HTML file preview (FilePreview.svelte, serveUrl branch) rendered served HTML in an iframe that hardcoded allow-same-origin. The terminal proxy serves the file root-relative (same origin as the app) and injects no CSP, and there is no default global CSP, so script in a previewed HTML file executed in the application's own origin and could read localStorage (the session token), enabling account takeover and, for admin or workspace.functions victims, server-side RCE via Functions. The sibling srcdoc branch already gates allow-same-origin behind the iframeSandboxAllowSameOrigin setting (off by default) and injects a CSP; the serveUrl branch never received that defense. Gate allow-same-origin on the serveUrl branch identically, so by default the preview runs at an opaque origin and its scripts cannot reach the parent context. Legitimate HTML preview rendering is unaffected.
Co-authored-by: manus-use <manus-use@users.noreply.github.com>
Enabling ENABLE_MILVUS_MULTITENANCY_MODE with the default MILVUS_URI (embedded Milvus Lite at DATA_DIR/vector_db/milvus.db) fails on the first embedding write: _create_shared_collection calls collection.create_index(RESOURCE_ID_FIELD) with no index params. A Milvus server auto-selects a scalar index type in that case, but Milvus Lite rejects the call with "create_index missing required 'index_type' parameter", so shared collection creation raises and every embedding write 500s (memory add, file upload, knowledge writes).
Keep the parameterless call as the first attempt so behavior on Milvus servers is unchanged, fall back to an explicit INVERTED scalar index, and if that also fails log a warning and continue. The scalar index only accelerates resource_id filters; inserts and filtered queries work without it, so a missing index must not break collection creation.
Verified against embedded Milvus Lite: shared collections now create (with the warning), and memory add, file upload and memory query succeed end to end. Against a Milvus server the first attempt is identical to the current code, so nothing changes where it works today.
SecurityHeadersMiddleware was the last middleware in the stack still
subclassing BaseHTTPMiddleware, after CommitSession, AuthToken,
WebsocketUpgradeGuard and Redirect were all moved to pure ASGI in
utils/asgi_middleware.py. BaseHTTPMiddleware re-buffers the response
body through an anyio task group, which has known issues with
streaming and Content-Length-bearing responses (e.g. the FileResponse
returned by /api/v1/audio/speech).
Reimplement it as a pure-ASGI middleware that stamps the configured
security headers onto the http.response.start message via
MutableHeaders and forwards all body chunks untouched, matching the
pattern already used by its four siblings. set_security_headers() and
all its helpers are unchanged.
Co-authored-by: classic298 <classic298@users.noreply.github.com>
The sidebar time-ago indicator rendered `created_at`, so the relative time stayed pinned to the chat's creation age and never reflected new activity. After sending a message the chat would jump to the top of the list (which sorts by `updated_at`) while still showing a stale label such as "3w", which is confusing.
The indicator was originally added using `updated_at` and was inadvertently switched to `created_at` during a later refactor. Restore `updated_at` (falling back to `created_at` when absent) so the timestamp matches the list ordering and updates whenever a chat is modified.
Fixes#26451
Translate all remaining untranslated strings in the German locale: 228 empty
values plus the genuinely translatable members of the value-equals-key set (240
entries total). Covers webhooks/events, external knowledge sources, context
compaction, terminal servers, OAuth/OIDC claims, skills, and directory
management.
Translations follow the existing locale conventions: formal "Sie" address,
established German terms (Websuche, Wissen, Werkzeuge, Wissensspeicher), and
brand/proper nouns, format names, and identical-in-German words left untouched.
Adds a {{USER_AGENT}} custom-header placeholder that relays the inbound
client's User-Agent to upstream model backends, so providers see the real
client instead of Open WebUI's internal aiohttp UA. This makes upstream
usage/cost attribution and backend telemetry possible, and is opt-in
per-connection (no global flag): admins add {{USER_AGENT}} to a connection's
custom headers in Admin > Settings > Connections.
The placeholder is sourced from the live inbound request (with a metadata
fallback for detached RAG/tool calls), so it resolves on every prompt-sending
path, not just chat completions:
- OpenAI completions, Responses API, and proxy — all route through
get_headers_and_cookies, which now passes the request into get_custom_headers.
- Anthropic Messages API (/api/v1/messages) — already covered, it delegates
to the chat completion handler.
- Ollama (/api/chat, /v1/completions, /v1/chat/completions, /v1/messages,
/v1/responses) — previously had no custom-header support at all; send_request
now applies per-connection custom headers (with templating) for every
Ollama prompt endpoint.
Custom headers are applied after the built-in user-info headers so explicit
admin-configured headers take precedence. The other existing placeholders
({{CHAT_ID}}, {{USER_ID}}, ...) now also work on the newly covered paths.
Frontend: the connection editor's Headers field is now shown for Ollama
connections too (previously gated to non-Ollama), so the placeholder can be
configured there.
Ref: open-webui/open-webui#26159
The base model selector in the model creation workspace listed every
model from the global store, ignoring the per-model visibility (hidden)
flag. Models marked as not visible were correctly hidden from the chat
model selector but still appeared in the workspace 'Base Model (From)'
dropdown.
Filter out hidden models (info.meta.hidden) to match the chat selector
behavior, while still keeping a model visible if it is the currently
selected base, so editing a preset built on a now-hidden base does not
show a blank dropdown.
Fixes#25665https://claude.ai/code/session_01CUvzYN9DjvwTpNCT5QL33B
Co-authored-by: Claude <noreply@anthropic.com>
routers/openai.py and routers/ollama.py decorated get_all_models with
`@cached(key=lambda _, user: ...)`. In aiocache 0.12, `key=` is a STATIC
cache key: get_cache_key returns `self.key` verbatim (the lambda object)
without calling it, so every caller collides to ONE shared entry within
the TTL. The intended per-user namespacing never happened — one user's
permission-filtered model list could be served to another user (or an
anonymous caller) during the cache window.
The per-call hook is `key_builder=` (called as key_builder(func, *args,
**kwargs)). Switch both sites to key_builder with a (func, request,
user=None) signature so the key is built per call. user=None mirrors
ollama's optional-user signature and stays correct whether user is passed
positionally, as a kwarg, or omitted.
Verified offline: old form returns the same key object for distinct users
(collision); new form yields distinct openai_all_models_<id> /
ollama_all_models_<id> keys, and the unauthenticated base key when no user.
These two were the only @cached(key=lambda ...) sites in the backend.
GET /api/v1/channels/{id}/messages/{message_id}/thread authorized only the URL channel, but get_messages_by_parent_id() appended the thread parent (loaded by id) without checking it belonged to that channel, so a caller could read a message from a channel they cannot access by passing its id as the thread root. Require the parent to be in the requested channel before returning it, and reject a posted parent_id/reply_to_id that does not belong to the channel.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A trailing space in the Open Terminal bearer token broke only the
interactive terminal: HTTP calls put the token in the Authorization
header, where the spec strips trailing whitespace, so the connection
test, file listing and tool calls all worked. The terminal WebSocket
can't set headers from the browser, so it sends the token inside a JSON
auth message that preserves the space verbatim, failing auth with
[Connection closed]. Normalize the key on save so whitespace never
enters storage, and trim it in the WebSocket auth message so existing
saved configs work without re-saving.
Fixes#25613
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Web search results stored as a vector collection were silently dropped
before retrieval when BYPASS_RETRIEVAL_ACCESS_CONTROL is False (the
default). The server-generated web_search file item carries a
'collection_name' but its 'web_search' type is not matched by any
explicit dispatch branch in get_sources_from_items, so it fell through
to the untrusted client-supplied collection_name branch and was ignored.
Add an explicit branch for type == 'web_search' items so the collection
is queried again. Access control is preserved: the collection still
passes through filter_accessible_collections, which already allowlists
web-search-* and bypasses only for admins.
Regression introduced when the retrieval access-control hardening gated
the bare collection_name fallback behind BYPASS_RETRIEVAL_ACCESS_CONTROL.
get_embedding_function runs at import time (main.py), so the empty-engine
+ no-model guard added in 55ca719b raised ValueError during startup and
bricked the instance: a blank embedding model set via the UI made the app
unbootable, with no way back into settings to undo it. Move the check into
the returned coroutine so construction always succeeds and the error only
fires when something actually embeds, surfacing as a clear request error
instead of the old cryptic 'NoneType has no attribute encode'.
Fixes#25634Fixes#25165
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The Sign Out button on the Account Pending overlay only cleared the
local token and redirected to /auth, skipping the backend signout. This
left the OAuth/OIDC session alive at the IdP, so re-login bounced the
user straight back into pending and never re-fetched updated role/group
claims. Call userSignOut() and honor the returned redirect_url, matching
the sidebar logout flow.
Fixes#25644
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The file picker and Files modal only render filenames and metadata, but
searchFiles()/getFiles() never passed content=false, so the backend
serialized data.content (full extracted text) of every file into each
response. On instances with many large documents this turns a metadata
list into tens of MB per request, re-issued on every picker keystroke.
The backend already supports content=false on GET /files/ and
GET /files/search; this just makes the list/search callers opt out of the
payload by default. The param stays overridable for callers that need it.
Fixes#25741
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A chat created in place (new chat -> first message) keeps chatIdProp
undefined because the URL is swapped via history.replaceState without a
remount, so none of the chatIdProp-gated updateLastReadAt paths fire and
the chat is never persisted with a last_read_at. Starting another new
chat via initNewChat reset $chatId without marking the outgoing chat
read, so on refresh updated_at > last_read_at (NULL) and the chat shows
as unread in the sidebar.
Mark the outgoing chat read in initNewChat, mirroring navigateHandler.
Fixes#25108
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
get_event_call() routed execute:python / execute:tool events to a client-supplied session_id after only checking the session was connected, not that it belonged to the requester. Verify the target session is owned by the requesting user (metadata user_id) before delivering, so a client cannot route code/tool execution into another user's session.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The Azure TTS handler (_tts_azure) interpolated the user-supplied voice,
and the locale derived from it, into the SSML xml:lang and <voice name>
attributes without XML-escaping, while the text body was already escaped
(2e75c6dbd). Escape both attributes too, so every user-derived value in
the SSML document is consistently encoded.
Co-authored-by: alanturing881 <alanturing881@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The web-ingest probe in get_content_from_url validated the URL at resolve
time (validate_url) but then fetched with a bare requests.get, which
re-resolves the hostname at connect time. An attacker-controlled name
server can answer with a public IP during validation and an internal IP
at connect, reaching cloud metadata / loopback / internal services (blind
always; binary content-types are read back to the caller). The
connection-layer guard (#24759) that closes this for the SafeWebBaseLoader
path was never mounted on this probe.
Route the probe through the same _SSRFSafeAdapter the loader uses, so the
resolution that feeds the TCP connect is re-validated against the
global-IP check.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
has_collection did `collection_name in self.client.list_collections()`,
but chromadb's list_collections() returns Collection objects (1.x), not
name strings — so the membership test is always False, even when the
collection exists. Compare against the collection names instead (with a
hasattr guard tolerating versions that yield plain names).
Found via the dependency-contract test suite (unit/deps/test_chromadb.py).
* Update SECURITY.md
* Update SECURITY.md
* Update SECURITY.md
* Update SECURITY.md
* Update SECURITY.md
* Update SECURITY.md
* Update SECURITY.md
* Update SECURITY.md
* Extend the already-fixed/monitoring rule to public PRs and credit
Broaden the rule from "already fixed" to also cover issues already being fixed in
the open (e.g. an open pull request), extend the commit-monitoring pattern to PRs,
and fold in the credit consequence on provable grounds rather than an unprovable
bad-faith claim: a report of an already-public, already-fixed-or-being-fixed issue
filed strictly last is a duplicate we cannot distinguish from scraping, so it earns
no advisory. Credit belongs to whoever found or fixed it, who forfeits it by
disclosing publicly instead of reporting confidentially first — so a public fix
earns no advisory and no credit for anyone.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Remove Rule 14 (One Vulnerability Per Report)
The one-CVE-per-vulnerability constraint it restated is a CVE Program counting
rule, already binding through the "Alignment with the CVE Program" section.
Dropping the standalone rule removes the duplication; bundled reports are still
split on that basis when they arise.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* security policy: surface "What a Valid Report Gets You" near the top, refresh date
Move the "What a Valid Report Gets You" section up to directly under the good-faith
reporting section (it leads with what reporters receive, rather than burying it
below the rules), and update the last-updated date.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Update SECURITY.md
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
_upload_file_async built its multipart body with
`aiohttp.streams.FilePayload(...)`, which no longer exists in aiohttp
(payload classes live in aiohttp.payload, and there is no FilePayload).
The reference sits inside a lazy closure, so import succeeds and only the
async Mistral OCR file-upload path blows up at runtime with
AttributeError on every call.
Mirror the working sync path: open the file in a context manager and let
MultipartWriter.append(file_obj, {...}) build a streaming
BufferedReaderPayload, with the POST issued inside the open() block so
the handle stays valid for the whole upload. Preserves the streaming /
memory-efficiency intent.
Found via the dependency-contract test suite (unit/deps/test_aiohttp.py),
which pins aiohttp.streams.FilePayload as absent.
constants.py subclassed opentelemetry.semconv.trace.SpanAttributes, which
is deprecated since semconv 1.25.0 (emits a DeprecationWarning; the pinned
0.63b1 has it live). Source the legacy span-attribute keys from the
non-deprecated _incubating attribute modules instead:
http.url / http.method / http.status_code <- _incubating http_attributes
db.name / db.statement / db.operation <- _incubating db_attributes
The stable http module renamed these (http.request.method, ...), so only
the incubating module preserves the original values. Custom keys
(db.instance/type/ip/port, error.*, result.*) stay literals.
Verified: emitted attribute keys are byte-identical before/after for every
key instrumentors.py reads, and importing constants no longer emits a
DeprecationWarning.
pydub imports the stdlib `audioop`, removed in Python 3.13, so audio
preprocessing would break there. requires-python is already capped at
< 3.13; this one-line pointer flags what to handle (audioop-lts, or drop
pydub) before raising that cap.
* chore: bump Python backend dependencies, drop unused peewee
Minor/patch + reviewed major bumps across requirements.txt,
requirements-min.txt, pyproject.toml and uv.lock; playwright image bumped in
docker-compose.playwright.yaml. peewee/peewee-migrate removed (zero imports).
Security-relevant: cryptography 46->48, authlib 1.6.10->1.7.2, PyJWT 2.11->2.13,
requests 2.33.1->2.34.2, RestrictedPython 8.1->8.2, pillow 12.1.1->12.2.0.
Reviewed majors: redis 7->8, pymilvus ->2.6.14, azure-search-documents 11->12,
chardet 5->7, unstructured 0.18->0.22, pycrdt 0.12->0.13.
Testing:
- Resolution: `uv lock` resolves the full bumped set with no conflicts; uv.lock
regenerated to match (peewee dropped, every pin including
azure-search-documents==12.0.0 resolves).
- Per-dependency contract tests (external tests repo, unit/deps/): 105 files,
2205 passed / 6 skipped, ruff-clean. One file per dependency pins the symbols,
signatures and behaviour the backend actually uses, so an API removal/rename in
a bumped version fails loudly instead of at runtime. Offline/deterministic.
- End-to-end embed->retrieve test driving transformers + sentence-transformers +
chromadb together through Open WebUI's real RAG path (cached model, in-memory
chroma, semantic retrieval asserted).
- Install/startup/health resolution gate added to the dep-bump workflow and the
integration suite (uv/pip resolve + uvicorn /health + Playwright dev visibility).
- Bugs surfaced while testing each got an isolated fix branch + regression test:
Mistral OCR aiohttp FilePayload (#25779), chroma has_collection (#25780),
aiocache per-user model-cache key (security), otel semconv deprecation,
pydub/audioop <3.13 note.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Bump python-multipart 0.0.22 -> 0.0.27 (CVE-2026-42561, CVE-2026-40347)
0.0.22 is affected by two DoS CVEs in the multipart parser that
Starlette/FastAPI run for every multipart/form-data request, so any
authenticated user hitting an upload endpoint can trigger them:
- CVE-2026-42561: unbounded part-header count/size -> CPU exhaustion (fixed 0.0.27)
- CVE-2026-40347: large multipart preamble/epilogue DoS (fixed 0.0.26)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Add a security-reporting link to the issue chooser
With blank issues disabled, the New Issue chooser only offered Bug Report and
Feature Request, leaving security reporters no obvious route and nudging them
toward filing vulnerabilities as public issues. Add a contact_links entry that
points to the Security Policy (/security/policy), where the "Report a
vulnerability" button opens a private GitHub advisory — keeping security
reports out of public issues.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Update config.yml
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Replace the recursive parseJSONString with an equivalent iterative unwrap.
The recursion re-parsed its own already-parsed result until JSON.parse threw;
on scalar JSON values (e.g. "5" -> 5) that recursed until a stack overflow
which was then silently caught — wasted work on every complete tool-call
payload. The loop returns the identical value in all cases (verified
byte-identical across 29 inputs incl. double/triple-encoded and partial JSON)
without the stack churn.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
has_access_to_file() derives file access from the objects a file is attached to
(knowledge bases, workspace models). Those branches returned True for any access_type
whenever the user held that permission on the object, write/delete included. Since a
user can create their own KB or model and attach any file they can merely READ (KB
attach and the model meta.knowledge validator both gate on read access only), a user
with read access to a victim file could launder it into write/delete: attach it to an
object they own, then rename, overwrite or delete it via the write-gated file routes
(POST /files/{id}/rename, /data/content/update, DELETE /files/{id}). This is the
residual of GHSA-vjqm-6gcc-62cr (CVE-2026-54012) left open by the read-only attach
validator (CWE-863).
An object now confers write/delete on a file only when the object's owner owns that
file, so delegation originates from the file's own owner. Read is unchanged (RAG and
shared-object reads still work), and legitimate delegation is preserved: a write grant
on an object whose owner owns the attached file still confers write. Applied to all
three object branches: knowledge base, file home collection, and workspace model.
Co-authored-by: rexpository <30176934+rexpository@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The direct image-edit route was the only image-edit surface with no authorization: it ran
on get_verified_user alone, while POST /generations enforces ENABLE_IMAGE_GENERATION +
features.image_generation and the built-in edit_image tool enforces ENABLE_IMAGE_EDIT +
features.image_generation. A verified non-admin user could therefore reach the configured
image-edit provider (spending IMAGES_EDIT_OPENAI_API_KEY) even when the administrator had
globally disabled image editing (ENABLE_IMAGE_EDIT=False) or denied the user image
generation.
Split the route from the shared impl (mirroring generate_images/image_generations): the new
/edit wrapper enforces ENABLE_IMAGE_EDIT and the per-user image-generation permission, then
delegates to image_edits(). The internal callers (the edit_image tool and the chat
middleware) already gate themselves and call image_edits() directly, so they are unaffected.
Co-authored-by: jagstack <52110932+jagstack@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
ws_terminal() interpolated the path parameter session_id directly into the upstream
terminal WebSocket URL and then appended ?user_id=<caller>, with no encoding or
validation (the HTTP sibling proxy_terminal runs _sanitize_proxy_path; this path ran
nothing). An encoded '?'/'&' smuggled through session_id survives Open WebUI's single
decode and is re-decoded by the upstream, injecting an attacker-chosen user_id ahead
of the appended one. Query parsing binds the first occurrence, so the orchestrator
resolves the spoofed user's terminal scope, letting a normal authenticated user
present another user's identity to the upstream (CWE-116/863).
Encode session_id as an opaque path segment with urllib.parse.quote(session_id,
safe=''). This neutralises '?'/'#'/'&' at any decode depth (the upstream's single
decode reverses only the quote, leaving the original delimiters inert as path
content), while legitimate UUID session ids pass through unchanged. The appended
user_id is then the only query parameter the upstream binds.
The separate concern that the forwarded identity is a bearer claim with no integrity
binding spans Open WebUI and the upstream terminal server and is not addressed here.
Co-authored-by: rexpository <30176934+rexpository@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
These were the only two ydoc Socket.IO handlers missing the SESSION_POOL guard that
every sibling (join/state/update) enforces. With always_connect=True admitting
unauthenticated sockets, a client that knew a note's document_id could broadcast
spoofed awareness (cursors/presence) and fake ydoc:user:left events into a live
editing room, attributed to any client-supplied user_id.
Both handlers now require an authenticated SESSION_POOL session; the awareness handler
additionally requires prior ydoc:document:join (room membership); and the broadcast
user_id is fixed to the authenticated identity instead of the client-supplied value,
removing the impersonation. Document content was never reachable (ydoc:document:update
already enforces write permission).
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The SCIM update_user (PUT) and patch_user (PATCH) handlers mapped the SCIM active field
unconditionally onto the role column (role = 'user' if active else 'pending'), so a routine
IdP sync or a misconfigured IdP that marked a locally-provisioned admin inactive would
silently strip that admin's role and could lock an instance out of its own administration.
Gate both active->role assignments on user.role != 'admin' so SCIM provisioning can
activate/deactivate ordinary users but never demotes an existing admin; admin role changes
continue to go through the dedicated admin endpoints. SCIM already cannot promote to admin
(active only maps to user/pending), so this is symmetric.
Credit to @HOHK0923 for surfacing the admin-demotion footgun.
Co-authored-by: HOHK0923 <118590749+HOHK0923@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
is_string_allowed does endswith() matching and was called with the full URL
(retrieval/web/utils.py) against WEB_FETCH_FILTER_LIST, so a blocklisted host with any
path (https://blocked.example/x) ended with /x, not the host, and slipped through; the
allowlist direction false-rejected legitimate URLs and admitted attacker URLs ending in
an allowed string. The same endswith caused label confusion at the hostname call site
(retrieval/web/main.py): corp.com matched evilcorp.com, 10.0.0.1 matched 110.0.0.1.
Add is_host_allowed(host, ...) matching on DNS label boundaries (host == pattern or
host.endswith('.' + pattern)), called with the parsed hostname at both web-fetch call
sites. is_string_allowed is left unchanged for the unrelated function-name filters
(utils/middleware.py, utils/tools.py).
The separate is_global guard (validate_url / _ssrf_safe_new_conn, active when
ENABLE_RAG_LOCAL_WEB_FETCH is off) already blocks RFC1918/loopback/link-local, so this
restores the admin's intended blocking of specific public hosts.
Co-authored-by: addcontent <59762500+addcontent@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The connection-layer DNS-rebinding guard (_SSRFSafeResolver / _SSRFSafeAdapter, PR #24759) was
mounted only on SafeWebBaseLoader. Two user-reachable image fetches validate the URL then fetch
it through the shared get_session() pool with the default resolver, so a TTL-0 rebinding answer
that passed validate_url reaches an internal address at connect:
- get_image_base64_from_url (utils/files.py): user image_url on every chat completion.
- load_url_image (routers/images.py, POST /api/v1/images/edit): user-supplied image field.
Add get_ssrf_safe_session() (a one-off aiohttp session mounting _SSRFSafeResolver) and use it
for both fetches, so the connect-time IP is re-validated and a rebound loopback / RFC1918 /
metadata address is rejected. The shared pool is left untouched for the admin-configured
image-generation callers, which legitimately reach internal hosts.
Co-authored-by: dhyabi2 <32069256+dhyabi2@users.noreply.github.com>
Co-authored-by: geo-chen <2404584+geo-chen@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
process_uploaded_file auto-links an uploaded file to the knowledge base named in
client-supplied metadata.knowledge_id, but it called Knowledges.add_file_to_knowledge_by_id
directly with no authorization, while the dedicated POST /knowledge/{id}/file/add endpoint
gates the same operation on owner / admin / write grant. A read-only collaborator (or any
verified user who knows the KB id) could therefore attach arbitrary files to a knowledge base
they cannot write: the KnowledgeFile membership row was committed, and the later vector-index
write check failed closed but left the row in place.
Gate the auto-link on the same owner/admin/write check before inserting the membership row;
otherwise skip the link and log. The file still uploads as the user's own file, it just is
not linked to a KB they cannot write.
Co-authored-by: jagstack <52110932+jagstack@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
generate_chat_completion() checks model access on line 200 only when bypass_filter is
False. When an arena model reaches this function without a pre-resolved selected_model_id,
which is the task and background path (the /api/v1/tasks/* endpoints call
generate_chat_completion directly rather than through process_chat_payload), the fallback
resolves the arena to an underlying model and recurses with bypass_filter=True, so the
resolved model's access check is skipped. An authenticated user with access to an arena
could therefore reach a model they are denied directly, and for the default or exclude
arena, whose candidate pool is every non-arena model, any model on the instance (CWE-862).
The normal chat path resolves the arena in process_chat_payload before this function, so its
resolved model is checked on line 200; the task path was not, which is the inconsistency.
Enforce check_model_access() on the resolved model in the fallback, before the
bypass_filter=True recursion, mirroring the normal-path check. Admins and already-bypassed
recursive calls are unaffected, and legitimate arena use of accessible models is unchanged.
Co-authored-by: rexpository <30176934+rexpository@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two lifecycle/authorization gaps let a deactivated (pending) account keep acting:
1. The background automation scheduler (execute_automation) rehydrated the owner by ID and
dispatched the chat pipeline without re-checking the owner. A user later set to pending,
or one whose features.automations permission was revoked, kept running scheduled
automations on the operator's provider credentials, even though the HTTP create/update/run
routes already gate on get_verified_user + features.automations. Re-gate the rehydrated
owner before dispatch: require role user/admin and, for non-admins, the features.automations
permission; otherwise record an error and skip the run.
2. check_model_access enforced per-model ACLs only for exactly role == 'user', so any other
non-admin role (a pending principal) fell through and was granted access. Enforce for every
non-admin role (admins still bypass), so the check fails closed (CWE-862, CWE-863).
Co-authored-by: rexpository <30176934+rexpository@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The decode-until-stable loops in _sanitize_proxy_path (terminals.py, cap 8) and
_safe_static_redirect_path (models.py, cap 2) proceed with whatever remains after the
cap instead of rejecting it. A path encoded more times than the cap therefore exits the
loop still percent-encoded, passes the literal '..' / prefix checks (the dots are still
%2E, not '..'), and is forwarded to the upstream terminal server, or emitted as a
redirect Location, which then decodes it once more and resolves the traversal. This is
the residual of the decode-until-stable hardening added for CVE-2026-54017: the cap is a
fixed depth, not a true stability guarantee.
Reject when the value is still not stable after the cap (unquote(x) != x), so anything
encoded more deeply than the cap fails closed rather than being forwarded. Legitimate
paths stabilize within a pass or two and are unaffected.
Co-authored-by: DavidCarliez <271374756+DavidCarliez@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix: block private-IP webhook URLs to close SSRF on caller-controlled URL
post_webhook(url, ...) in utils/webhook.py forwards the URL straight to
aiohttp.ClientSession.post with no SSRF gate. The URL is caller-controlled
on two surfaces:
- User notification settings under ENABLE_USER_WEBHOOKS=true — any
authenticated user can set the URL their notifications POST to.
- Automation notification triggers (calendar alerts, etc.).
Without a gate, the URL can target cloud metadata (169.254.169.254 /
fd00:ec2::254), localhost-bound services, RFC1918 internal hosts, or any
other private address reachable from the server process. Blind SSRF — no
response body returned to the caller — but enough to enumerate internal
services via response timing / status codes, and on cloud deployments
enough to issue requests against IMDSv1 if available.
Call validate_url() at the top of post_webhook. The function blocks
private/reserved IPs when ENABLE_RAG_LOCAL_WEB_FETCH is False (the
default), is the project's chosen SSRF gate, and is already applied to
the equivalent fetch surfaces (retrieval, image-load, OAuth profile
picture). Operators who legitimately need to webhook to private IPs
(internal monitoring, self-hosted Slack alternatives, etc.) can set
ENABLE_RAG_LOCAL_WEB_FETCH=True — same opt-out as the other gated
surfaces.
Scope intentionally limited to webhooks. The OAuth discovery and
external reranker paths cwanglab also flagged are admin-configured with
intentional private-IP defaults (reranker defaults to
http://localhost:8080/v1/rerank) and are out of scope per Rule 9 — the
admin owns the URL choice and the operator opt-out exists for them too.
Reported by cwanglab in GHSA-5x9f-85cg-w3hf (cluster canonical with six
closed siblings: g36v-23gj-j69x, 6j8f-h58v-xgmw, xpwv-52pm-p8hj,
v9gp-hv2c-9qv8, fw7w-jrw7-p3v9, x7xq-74rg-m8mf).
Co-authored-by: cwanglab <cwanglab@users.noreply.github.com>
* fix: also pass allow_redirects=False on webhook post_webhook session.post
Companion to the previous commit. validate_url() only validates the
initial URL; aiohttp's default allow_redirects=True would still follow
a 302 to a private-IP target. Same redirect-bypass class as the rh5x
cluster's five call sites, sixth call site to receive the same gate.
Co-authored-by: cwanglab <cwanglab@users.noreply.github.com>
---------
Co-authored-by: cwanglab <cwanglab@users.noreply.github.com>
The Qdrant backends implemented delete(ids=...) as a payload filter on
metadata.id, but points are stored with the item id as the Qdrant point id
(see _create_points), and not every point carries an id in its payload.
Memory points store only {created_at} in metadata (KB metadata embeddings
likewise), so deleting a single memory matched nothing and left an orphaned
vector that kept being injected into RAG context.
Delete by point id instead: PointIdsList for the standard backend, and a
tenant-scoped HasIdCondition for multitenancy (point ids are unique, so tenant
isolation is preserved). Filter-based deletion is unchanged.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix: gate chat_completion channel: branch on channel access + message scoping
When chat_id starts with 'channel:' the chat-completion handler skips
the chat ownership / storage block below it. Nothing replaced that
gate. The downstream channel emitter in socket/main.py:_make_channel_
emitter writes to Messages.update_message_by_id using a caller-supplied
message_id pulled from form_data['id'], with no membership check, no
write-access check, and no validation that the message_id belongs to
the channel.
Net effect: any authenticated user could submit
chat_id='channel:<any-channel-uuid>' + id='<any-message-uuid>' and
overwrite that message with attacker-controlled LLM output. Cross-
channel writes worked too — private channels, DMs, channels the
caller has no access to. Original author attribution stayed intact on
the overwritten row.
Add the missing checks at the channel: branch:
1. Channel must exist (404 otherwise).
2. Non-admin caller must have write access to the channel — membership
for group/dm channels, AccessGrants permission='write' for others.
3. The message_id (if supplied) must belong to the same channel — a
caller with write access to channel A cannot use this path to
overwrite a message in channel B.
Behaviour change is limited to callers who were exploiting the gap:
legitimate flows that supply a message_id under their own channel
membership continue to work unchanged.
Co-authored-by: sfwani <sfwani@users.noreply.github.com>
* chore: trim verbose comment on channel: branch gate
---------
Co-authored-by: sfwani <sfwani@users.noreply.github.com>
The history diff, delete, and version-restore routes authorize the URL
prompt_id but then act on a caller-supplied history/version id without
checking it belongs to that prompt (IDOR). Filter by prompt_id in
compute_diff and delete_history_entry, and reject a cross-prompt version_id
in update_prompt_version.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Firecrawl /search returns either `{"data": [...]}` (flat list — v1, and
what frost19k reported on #23966) or `{"data": {"web": [...]}}` (v2,
current production). The parser only handled the dict shape:
data = response.get('data') or {}
results = data.get('web') or []
On a list-shape response, `data.get('web')` raised AttributeError,
caught by the function's outer try/except, and `search_firecrawl`
silently returned []. Web search worked against v2 endpoints but is one
upstream-format-change away from failing closed again. Accept either.
The native function-calling tool resolver in utils/tools.py applies five
gates before exposing execute_code as a builtin tool: builtin-category
enable, ENABLE_CODE_INTERPRETER global config, model capability,
features.code_interpreter request flag, and the per-user
features.code_interpreter permission.
The legacy XML-tag detection path in streaming_chat_response_handler
applied only the request-flag gate. Brings the legacy path to parity by
running the same five-gate check before activating tag detection.
Behaviour change is limited to deployments that previously relied on
the asymmetry — admins who set ENABLE_CODE_INTERPRETER=False or revoked
the per-user permission, on the legacy tool-calling mode, with the
client supplying features.code_interpreter=true. Any of those three
conditions met now correctly disables tag detection.
Co-authored-by: sfwani <sfwani@users.noreply.github.com>
Open WebUI's collection ACL accepted any unknown name as a
legacy/ephemeral collection. In Milvus multi-tenancy mode that name
becomes the `resource_id` and is interpolated unescaped into a SQL-like
Milvus expression — `resource_id == '<name>'` — so a name like
x' or resource_id != '' or resource_id == 'x
turns the filter into a tautology and returns every tenant's chunks
from the shared collection.
All collection names Open WebUI generates are UUIDs, SHA-256 hex
digests, or fixed-prefix variants of those — they all fit
[A-Za-z0-9_-]. Add a strict format check in
filter_accessible_collections (utils.py) that drops any name outside
that set before any ACL or vector-store lookup, applied even on the
admin bypass path. _validate_collection_access then surfaces the dropped
name as a 403.
As defense in depth, MilvusClient now validates resource_id at every
expression-construction site and escapes single quotes / backslashes in
any other string interpolated into a filter (delete ids, metadata
filter values). Non-string filter values are typed-checked instead of
str()-formatted.
Co-authored-by: Claude <noreply@anthropic.com>
update_chat_by_id re-derived assistant content from `output` on every
save (serialize_output) so frontend edits to output items reflect in
content. But it ran unconditionally, so content set independently of
output — an __event_emitter__ {"type":"replace"} from an Action, or an
outlet filter footer — was reverted to the original output-derived text
on the next save. The reload reads chat.chat directly, so the change
vanished after navigating away (regression vs 0.9.2, which predates the
output mechanism).
Re-derive only when the message's `output` actually changed versus what's
stored, which still reflects genuine output edits but leaves
independently-set content intact.
Fixes#24585
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Symptom
On a fresh install (zero users) the frontend shows the mandatory
"Create Admin Account" onboarding screen, but POST
/api/v1/auths/signup returns 403 ACCESS_PROHIBITED ("You do not have
permission to access this resource."). Wiping the database does not
help when the config layer is backed by Redis (the value survives in
the Redis/valkey volume), or when only the user table is cleared (the
config row survives in Postgres). The instance is then unrecoverable
through the UI.
Root cause
signup_handler() auto-sets request.app.state.config.ENABLE_SIGNUP =
False immediately after the first admin is created. That value is
persisted by the config layer (the Postgres config table, and Redis
when REDIS_URL is set). On a later zero-user database the persisted
False is read back, so ENABLE_SIGNUP resolves False even though no
users exist. The old gate was:
if WEBUI_AUTH:
if not ENABLE_SIGNUP or not ENABLE_LOGIN_FORM:
if has_users or not ENABLE_INITIAL_ADMIN_SIGNUP:
403
ENABLE_INITIAL_ADMIN_SIGNUP defaults to False, so with zero users the
inner test (has_users or not ENABLE_INITIAL_ADMIN_SIGNUP) is True, and
a stale ENABLE_SIGNUP=False trips the outer test, producing a 403 on
the only UI path that can create the first admin. The frontend decides
to show onboarding purely from user_count == 0, so frontend and
backend disagree and the instance bricks.
Change
Split the gate by has_users. Subsequent signups (has_users True) are
unchanged: still gated by ENABLE_SIGNUP and ENABLE_LOGIN_FORM. The
first user (has_users False, the bootstrap admin the onboarding screen
invites) is gated only by the admin-chosen ENABLE_LOGIN_FORM (the
documented SSO-only hard-disable) unless ENABLE_INITIAL_ADMIN_SIGNUP is
set. It is no longer gated by ENABLE_SIGNUP, which in the zero-user
state is never an admin decision but the post-first-admin auto-disable
leaking across a database reset.
Why this is safe (full case analysis)
For WEBUI_AUTH the gate has 16 input combinations over (has_users,
ENABLE_SIGNUP, ENABLE_LOGIN_FORM, ENABLE_INITIAL_ADMIN_SIGNUP). Old and
new are identical in 15 of them:
* All 8 has_users=True cases: both reduce to "403 iff not
ENABLE_SIGNUP or not ENABLE_LOGIN_FORM". Unchanged.
* 7 of the 8 has_users=False cases: identical.
The only changed case is has_users=False, ENABLE_SIGNUP=False,
ENABLE_LOGIN_FORM=True, ENABLE_INITIAL_ADMIN_SIGNUP=False: old
behaviour 403, new behaviour allow. The new condition is a strict
subset of the old (new-403 implies old-403), so the change never newly
blocks any request that previously succeeded; it only stops blocking
that one bootstrap state.
That state has no legitimate deployment. With the login form enabled
and zero users the onboarding form is already served, and the only
operator-configurable way to keep the first signup closed (SSO-only:
ENABLE_LOGIN_FORM=False, optionally with ENABLE_INITIAL_ADMIN_SIGNUP)
is preserved byte for byte. ENABLE_SIGNUP=False with zero users is not
an operator choice, it is the automatic post-first-admin disable, so
the old behaviour there was purely a brick with no recovery path. No
security control is weakened: ENABLE_LOGIN_FORM and
ENABLE_INITIAL_ADMIN_SIGNUP keep their exact meaning, and the
WEBUI_AUTH=False path is untouched.
This is not Redis-specific: it reproduces with Redis disabled through
the Postgres config table alone (clear the user table, keep the config
row).
Verification
Drove the real signup endpoint across a 10-case matrix on freshly
migrated databases, including the full end-to-end first-admin creation
(returns role=admin, row persisted as admin) and the preserved
SSO-only, subsequent-signup and no-auth behaviours. All pass.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The startup banner uses Unicode box-drawing characters. On a stdout that
can't encode them (Windows cp1252, or redirected/headless/pythonw output)
print() raises UnicodeEncodeError and aborts startup. This blocks running
open-webui serve headless on Windows.
Guard the banner print and fall back to a plain ASCII line so startup
always proceeds regardless of the console encoding.
Fixes#24965
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
serve_cache_file gated the resolved path with file_path.startswith(os.path.abspath(CACHE_DIR)) without a trailing os.sep, so any path resolving to a sibling whose name starts with the cache-dir basename (e.g. cache_backup, cached_models) passed the prefix check. Authenticated users could read files from such siblings via /cache/../<sibling>/<file>. Appending os.sep to the prefix closes the bypass; deep traversal and absolute paths were already correctly blocked.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The proxy gathered _client_to_upstream and _upstream_to_client with
return_exceptions=True. When upstream sends a graceful CLOSE,
_upstream_to_client returns but gather keeps waiting on
_client_to_upstream, which is blocked in ws.receive() until the browser
disconnects. The handler stays pending and the finally: session.close()
cleanup is deferred, leaking a ClientSession and an open browser socket.
Use asyncio.wait(return_when=FIRST_COMPLETED) and cancel the pending
sibling, so the proxy unwinds as soon as either direction finishes. The
pumps' bare except Exception already lets CancelledError (a BaseException)
propagate, so cancellation is clean and they need no change.
Fixes#25464
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat: cap profile image data URI size to bound model/avatar bloat
validate_profile_image_url() validated data-URI format (MIME allowlist,
SVG rejection, scheme checks) but never its length, so a valid
data:image/...;base64,<huge> passed for both custom-model icons and user
avatars. Large inline images bloat Postgres and the Redis MODELS hash and
degrade model-list latency.
Add PROFILE_IMAGE_MAX_DATA_URI_SIZE (default 256 KiB, 0 disables) and
reject oversized data URIs in the shared validator, so both model meta
(ModelMeta.profile_image_url) and user avatars (UpdateProfileForm) are
bounded at one chokepoint. ModelMeta already clears invalid values to
None on read, so existing oversized icons stop propagating into the
MODELS hash on the next refresh.
Fixes#25468
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix: default PROFILE_IMAGE_MAX_DATA_URI_SIZE to None (no cap)
Per review: opt-in rather than a 256 KiB default. Unset leaves data URIs
uncapped; the validator already skips the check on a falsy value.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
getChatEventEmitter starts a setInterval emitting a `usage` socket event
every second; clearInterval ran only after sendMessageSocket resolved, so
any throw/reject left the interval firing for the page lifetime. Each
failed send added another orphaned interval, inflating server-side usage
accounting and growing CPU/memory over a session.
Wrap the send in try/finally so the interval is always cleared, on both
the happy path and any thrown/rejected path. The exception still
propagates unchanged.
Fixes#25465
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
_sanitize_proxy_path decoded the proxy path once before the '..' check, so a double-encoded payload (%252e%252e) survived the check as %2e%2e and was then re-decoded into '..' by the upstream terminal server, defeating the traversal guard. Decode until stable so no encoded traversal sequence can reach the upstream. Single-encoded payloads were already rejected; this closes the double (and deeper) encoding bypass.
Co-authored-by: sermikr0 <230672901+sermikr0@users.noreply.github.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
bypass_system_prompt is an internal flag used by utils/middleware.py and utils/chat.py to skip applying the model system prompt on recursive base-model calls, but it was still declared as a positional argument on the openai/ollama chat-completion route handlers, so FastAPI bound it from the query string. Move it to request.state so external clients cannot set it, matching how bypass_filter is handled.
Drop the argument from both route signatures and read getattr(request.state, 'bypass_system_prompt', False); utils/chat.py sets request.state.bypass_system_prompt alongside bypass_filter and drops the kwarg from the two route-handler calls (the recursive self-calls keep it). Mirrors c0385f60b.
Co-authored-by: anishgirianish <161533316+anishgirianish@users.noreply.github.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The 't0p-s3cr3t' default was dead code on every supported startup path:
start.sh, start_windows.bat and `open-webui serve` all set or
auto-generate WEBUI_SECRET_KEY before the backend imports env.py. It was
only ever reachable by invoking uvicorn directly, which is unsupported
and unsafe (the app would then sign tokens/cookies with a public,
hardcoded key). It also keeps getting reported as a vulnerability because
it looks dangerous, even though it is unreachable in practice.
Drop the fallback (default to '') so an unset key is caught by the
existing WEBUI_AUTH guard, and replace the vague error with a clear,
actionable message explaining that the key is a hard requirement and how
the supported start methods provide it. Exit cleanly via SystemExit
instead of raising a ValueError traceback.
WEBUI_AUTH=False keeps working unchanged (key defaults to '').
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
renderMermaidDiagram returned raw mermaid SVG, which FilePreview.svelte injects
via wrapper.innerHTML = svg. Mermaid runs with securityLevel: 'loose', so it
neither sanitizes click hrefs (formatUrl skips sanitizeUrl) nor DOMPurifies its
output; a .md file with a click X href "javascript:..." directive (or an
HTML-label payload) therefore executes script in the app origin when previewed.
The chat path was already safe because SVGPanZoom DOMPurifies before rendering;
file preview was not.
Sanitize at the source: renderMermaidDiagram now returns DOMPurify-cleaned SVG
via a shared sanitizeSvg helper (same policy as SVGPanZoom), so every consumer
including the FilePreview innerHTML sink receives safe output.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
insert_chat_files() stored any caller-supplied file_id with no ownership
check, so a user could attach another user's file to their own chat and
then read it through the shared-chat access path in has_access_to_file().
Filter file_ids to those the caller owns, is admin for, or can read.
Also repairs an UnboundLocalError introduced in 260ead64d: the existing
duplicate-check referenced `session` before it was assigned (db=session),
so the function threw on every call and no chat_file rows were persisted.
ModelMeta.profile_image_url now runs validate_profile_image_url, rejecting SVG/script data URIs (matching UserUpdateForm and ChannelWebhookForm). The /model/profile/image endpoint enforces the PROFILE_IMAGE_ALLOWED_MIME_TYPES allowlist and sets X-Content-Type-Options: nosniff, so an SVG data URI can no longer be served inline on-origin. Closes the fourth profile-image XSS sink missed by the user and webhook fixes.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When called without attached model knowledge and given a caller-supplied knowledge_id, search_knowledge_files passed it straight to Knowledges.search_files_by_id, which does not enforce ownership on knowledge_id. An authenticated user who happened to know a target UUID could enumerate file metadata (filename, file id, KB id, KB name, updated_at) from any knowledge base, bypassing the AccessGrants permission model.
Mirror the same admin/owner/AccessGrants check the attached-KB branch already uses, matching the sibling query_knowledge_files function.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The lookup-prompt-by-command endpoint's only frontend wrapper,
getPromptByCommand, is dead: nothing imports or calls it, the path is
referenced nowhere else, and the route handler has no internal caller
(slash-command resolution happens client-side from the loaded prompt
list). Removes the route handler, its section header, and the dead
wrapper.
The shared Prompts.get_prompt_by_command data-layer method is kept: it
is still used by create/update prompt validation (prompts.py:175, 275,
340). PromptAccessResponse and AccessGrants are untouched.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Classic298