update-stop-gate
Pass
Audited by Gen Agent Trust Hub on Oct 3, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [COMMAND_EXECUTION]: The skill uses
python3 -c "import py_compile; ..."to verify the syntax of thestop-gate.pyfile after modification. This involves executing shell commands on the local system. - [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied examples of "bad behavior" to generate new rules.
- Ingestion points: User descriptions provided during the "Trigger phrases" phase.
- Boundary markers: None explicitly defined in the instructions for how the user input is delimited.
- Capability inventory: The skill performs file writes to
~/.claude/autonomy-rules.mdand~/.claude/hooks/stop-gate.py, and executes a Python compilation check via the shell. - Sanitization: The skill instructions include a manual "Scrub for internal/proprietary language" step, but there is no automated escaping or sanitization to prevent the agent from adopting malicious instructions embedded in the user's "example behavior."
- [METADATA_POISONING]: The
verification/commands-tested.mdfile contains a termination labelshipped_lite, which is a metadata field that could be used to deceive automated scanners or the agent itself regarding the skill's production status or safety.
Audit Metadata