Summary
Chat histories are stored as an unvalidated JSON object. After a message is deleted, the code that picks the chat's new current message walked down the childrenIds links without recording where it had already been. Any account with the default user role could store a chat whose messages list each other as children, then delete a message from it, and the walk would run forever. That walk runs on the server's request loop, so it blocks every other user's requests until the process is killed.
Preconditions
One account with the default user role. No administrator rights, no additional permissions, no configuration change and no non-default setting: creating and deleting chats is available to every user out of the box. The attack runs entirely against the attacker's own chat, so no knowledge of any other user's data is needed. Versions before 0.10.0 are unaffected because neither the message-deletion endpoint nor the affected code existed.
Impact
The walk is synchronous and runs on the asyncio event loop, so while it spins, every request from every user is blocked, including unauthenticated /health and administrator endpoints. External health checks and orchestrator liveness probes fail alongside the UI. This is a pure CPU pin with no list growth, so a worker consumes one full core with flat memory and has to be killed rather than being reclaimed by an out-of-memory kill. The work is not cancelled when the client disconnects, so a single fire-and-forget request is enough and the attacker can disconnect immediately. The malformed chat stays in the database, so the condition re-arms on the next deletion attempt against that chat. No data is disclosed, altered or deleted.
Fix
Fixed in 0.11.1 by open-webui/open-webui@b933292. The walk now records the ids it has already passed through, so it terminates after at most one step per stored message whatever the message contents are. Upgrading fully resolves the issue, including for chats stored while the deployment was on an affected version, which become harmless once the walk terminates.
Root cause
Affected component: the chat-history message deletion helper in backend/open_webui/models/chats.py, reached from DELETE /api/v1/chats/{id}/messages/{message_id}. Affected setup: all builds from 0.10.0 up to and including 0.11.0.
Resolving the chat's new current message after a deletion means descending to the deepest remaining child, and that descent had no notion of where it had already been, so two messages naming each other as children sent it back and forth indefinitely. The write path does not validate the structure of a chat's history, and the child-link rebuild that runs on other write paths does not apply when a chat is created, so a history whose childrenIds form a cycle was persisted exactly as submitted.
Proof of concept
As a default-role user on a running instance, store one chat of roughly 300 bytes whose messages reference each other as children:
{"chat":{"title":"poc","history":{"currentId":"C","messages":{
"A":{"id":"A","parentId":null,"role":"user","content":"a","childrenIds":["B","C"],"timestamp":1},
"B":{"id":"B","parentId":"A","role":"assistant","content":"b","childrenIds":["A"],"timestamp":2},
"C":{"id":"C","parentId":"A","role":"user","content":"c","childrenIds":[],"timestamp":3}}}}}POST /api/v1/chats/new stores it verbatim, with the child links unchanged. A single DELETE /api/v1/chats/{id}/messages/C as the same user then never returns:
unauthenticated GET /health -> HTTP 000 after 10.007s
DELETE -> HTTP 000 after 20.007s
Sampled during the hang, the worker had consumed 42.9 seconds of CPU on one fully occupied core with flat resident memory, and unauthenticated GET /health timed out for as long as the process lived, including after the attacker's connection closed.
On 0.11.1 the same payload returns HTTP 200 in 0.011s, /health stays available throughout, and deleting a middle message from a well-formed four-message chain still resolves the current message correctly.
Credits
@Classic298, who found the unguarded descent through the chat's child links, demonstrated the resulting server-wide outage end to end against a live instance, and supplied the fix.
References
Classic298 Reporter
Summary
Chat histories are stored as an unvalidated JSON object. After a message is deleted, the code that picks the chat's new current message walked down the
childrenIdslinks without recording where it had already been. Any account with the defaultuserrole could store a chat whose messages list each other as children, then delete a message from it, and the walk would run forever. That walk runs on the server's request loop, so it blocks every other user's requests until the process is killed.Preconditions
One account with the default
userrole. No administrator rights, no additional permissions, no configuration change and no non-default setting: creating and deleting chats is available to every user out of the box. The attack runs entirely against the attacker's own chat, so no knowledge of any other user's data is needed. Versions before 0.10.0 are unaffected because neither the message-deletion endpoint nor the affected code existed.Impact
The walk is synchronous and runs on the asyncio event loop, so while it spins, every request from every user is blocked, including unauthenticated
/healthand administrator endpoints. External health checks and orchestrator liveness probes fail alongside the UI. This is a pure CPU pin with no list growth, so a worker consumes one full core with flat memory and has to be killed rather than being reclaimed by an out-of-memory kill. The work is not cancelled when the client disconnects, so a single fire-and-forget request is enough and the attacker can disconnect immediately. The malformed chat stays in the database, so the condition re-arms on the next deletion attempt against that chat. No data is disclosed, altered or deleted.Fix
Fixed in 0.11.1 by open-webui/open-webui@b933292. The walk now records the ids it has already passed through, so it terminates after at most one step per stored message whatever the message contents are. Upgrading fully resolves the issue, including for chats stored while the deployment was on an affected version, which become harmless once the walk terminates.
Root cause
Affected component: the chat-history message deletion helper in
backend/open_webui/models/chats.py, reached fromDELETE /api/v1/chats/{id}/messages/{message_id}. Affected setup: all builds from 0.10.0 up to and including 0.11.0.Resolving the chat's new current message after a deletion means descending to the deepest remaining child, and that descent had no notion of where it had already been, so two messages naming each other as children sent it back and forth indefinitely. The write path does not validate the structure of a chat's history, and the child-link rebuild that runs on other write paths does not apply when a chat is created, so a history whose
childrenIdsform a cycle was persisted exactly as submitted.Proof of concept
As a default-role user on a running instance, store one chat of roughly 300 bytes whose messages reference each other as children:
{"chat":{"title":"poc","history":{"currentId":"C","messages":{ "A":{"id":"A","parentId":null,"role":"user","content":"a","childrenIds":["B","C"],"timestamp":1}, "B":{"id":"B","parentId":"A","role":"assistant","content":"b","childrenIds":["A"],"timestamp":2}, "C":{"id":"C","parentId":"A","role":"user","content":"c","childrenIds":[],"timestamp":3}}}}}POST /api/v1/chats/newstores it verbatim, with the child links unchanged. A singleDELETE /api/v1/chats/{id}/messages/Cas the same user then never returns:Sampled during the hang, the worker had consumed 42.9 seconds of CPU on one fully occupied core with flat resident memory, and unauthenticated
GET /healthtimed out for as long as the process lived, including after the attacker's connection closed.On 0.11.1 the same payload returns HTTP 200 in 0.011s,
/healthstays available throughout, and deleting a middle message from a well-formed four-message chain still resolves the current message correctly.Credits
@Classic298, who found the unguarded descent through the chat's child links, demonstrated the resulting server-wide outage end to end against a live instance, and supplied the fix.
References