Poco Ai
Monthly
Incomplete session cleanup in poco-ai/poco-agent up to version 0.5.4 can permanently destroy the host user's ~/.claude directory - including all Claude Code chat history stored under ~/.claude/projects - when the agent runs outside a container environment. The WorkspaceManager._setup_session_persistence method performs an unconditional rmtree followed by a symlink swap on ~/.claude without verifying a container context, and the corresponding cleanup phase can additionally unlink pre-existing ~/.claude symlinks unrelated to poco-agent. A publicly available exploit exists via GitHub issue #131 and PR #132, though the CVSS 4.0 base score of 2.9 reflects high attack complexity and impact limited to low integrity and availability loss with no confidentiality breach.
Server-side request forgery in poco-ai/poco-claw up to version 0.5.4 allows remote unauthenticated attackers to coerce the application server into issuing arbitrary HTTP requests to attacker-controlled destinations via the unvalidated callback_url parameter in the run_task API function. The flaw is exposed through the task execution endpoint at executor/app/api/v1/task.py and carries a publicly available proof-of-concept exploit referenced on GitHub. No vendor patch has been released; the associated GitHub issue was closed automatically due to inactivity, indicating no active maintenance response and indefinite risk exposure.
Missing authentication on the executor_manager control-plane API in poco-ai/poco-claw through version 0.5.4 allows adjacent-network attackers to invoke the `create_task` endpoint without any credentials, enabling unauthorized task submission to the AI agent execution engine. The CVSS 4.0 vector (AV:A) confirms exploitation requires adjacency to the internal network segment hosting the executor manager service - it is not directly internet-exploitable. Public exploit code has been disclosed via GitHub issues #136/#137 and PR #135; no CISA KEV listing exists at time of analysis.
Unauthenticated authorization bypass in poco-ai/poco-claw's Workspace API allows remote attackers to access arbitrary users' workspace files by manipulating the user_id parameter in the get_workspace_file function. All versions through 0.5.4 are affected, with the executor manager's control-plane routes completely lacking token-based authentication before the fix. Publicly available exploit code exists (POC confirmed via GitHub issue #133 and PR #135); no CISA KEV listing at time of analysis.
Incomplete session cleanup in poco-ai/poco-agent up to version 0.5.4 can permanently destroy the host user's ~/.claude directory - including all Claude Code chat history stored under ~/.claude/projects - when the agent runs outside a container environment. The WorkspaceManager._setup_session_persistence method performs an unconditional rmtree followed by a symlink swap on ~/.claude without verifying a container context, and the corresponding cleanup phase can additionally unlink pre-existing ~/.claude symlinks unrelated to poco-agent. A publicly available exploit exists via GitHub issue #131 and PR #132, though the CVSS 4.0 base score of 2.9 reflects high attack complexity and impact limited to low integrity and availability loss with no confidentiality breach.
Server-side request forgery in poco-ai/poco-claw up to version 0.5.4 allows remote unauthenticated attackers to coerce the application server into issuing arbitrary HTTP requests to attacker-controlled destinations via the unvalidated callback_url parameter in the run_task API function. The flaw is exposed through the task execution endpoint at executor/app/api/v1/task.py and carries a publicly available proof-of-concept exploit referenced on GitHub. No vendor patch has been released; the associated GitHub issue was closed automatically due to inactivity, indicating no active maintenance response and indefinite risk exposure.
Missing authentication on the executor_manager control-plane API in poco-ai/poco-claw through version 0.5.4 allows adjacent-network attackers to invoke the `create_task` endpoint without any credentials, enabling unauthorized task submission to the AI agent execution engine. The CVSS 4.0 vector (AV:A) confirms exploitation requires adjacency to the internal network segment hosting the executor manager service - it is not directly internet-exploitable. Public exploit code has been disclosed via GitHub issues #136/#137 and PR #135; no CISA KEV listing exists at time of analysis.
Unauthenticated authorization bypass in poco-ai/poco-claw's Workspace API allows remote attackers to access arbitrary users' workspace files by manipulating the user_id parameter in the get_workspace_file function. All versions through 0.5.4 are affected, with the executor manager's control-plane routes completely lacking token-based authentication before the fix. Publicly available exploit code exists (POC confirmed via GitHub issue #133 and PR #135); no CISA KEV listing at time of analysis.