agent_team

Warn

Audited by Gen Agent Trust Hub on Mar 11, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The deep_think engine utilizes asyncio.create_subprocess_exec to run Python scripts and testing commands (pytest) within a local sandbox directory. This involves executing arbitrary code generated by the agent, which can be influenced by user-provided task instructions.\n- [REMOTE_CODE_EXECUTION]: The dispatch_task and dispatch_batch_tasks tools facilitate sending instructions to other agent instances listening on local network ports (localhost). This mechanism allows an agent to trigger the execution of arbitrary tasks in other processes.\n- [PROMPT_INJECTION]: The skill documentation includes an 'Emergency Takeover' rule that explicitly directs the agent to 'switch identity' and utilize high-privilege local tools such as bash or file_editor if worker dispatch fails, potentially bypassing intended restrictions.\n- [DATA_EXFILTRATION]: The sync_task_context tool enables the agent to query and retrieve session data and conversation histories from other nodes on the local network, allowing for the movement of sensitive session information across different agent contexts.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 11, 2026, 02:34 AM