9 HIGH · 1 MED · 2 INFO · THREAT ORANGE · 12 items · Generated in 224s
Read-only users can modify a knowledge base by adding attacker-controlled files without write permission.
Revoked tokens can still open realtime sessions after logout, letting unauthorized access persist on websocket and terminal channels.
One authenticated message can freeze the default single-worker instance and deny service to all users until it is killed.
An authenticated user can read private channel messages from channels they should not access due to broken authorization checks.
An authenticated user can turn read-only access into write or delete access on another user's file.
Authenticated users can alter messages in private or DM channels they do not belong to, breaking integrity and trust.
Authenticated users can reach restricted models through task endpoints, breaking intended model access controls.
Shared caching can expose one user's permitted model list to another, breaking tenant isolation and confidentiality.
Shows open-source AI agents can automate post-exploitation, lowering attacker effort and speeding intrusions.
Shows an AI agent being used autonomously for post-exploitation and privilege hunting inside a government network.
Funding news signals market activity, not a direct security issue or exploitable weakness.
Signals rising patch volume from AI-assisted bug hunting, increasing defender workload without a specific AI exploit described.