The Agent Skills Directory

[COMMAND_EXECUTION]: Use of the --dangerously-skip-permissions flag when invoking the claude CLI. This bypasses the tool's built-in security sandbox and allows the agent to execute operations without user confirmation or permission checks.\n- [COMMAND_EXECUTION]: Vulnerability to shell command injection through the direct interpolation of user-provided task descriptions ([TASK]) into the CLI parameters without sanitization or escaping.\n- [EXTERNAL_DOWNLOADS]: Automated installation of Python packages using uv pip install when a task fails due to missing dependencies, which can be leveraged to install untrusted third-party code.\n- [COMMAND_EXECUTION]: Spawning a local background server using python dashboard_serve.py. This opens a local network port and executes a script that could be modified by prior steps.\n- [COMMAND_EXECUTION]: Potential path traversal or unauthorized directory creation through the use of an unsanitized task_name variable to define the TASK_DIR.

biomed-dispatch