maintain-formal-verification-skill
Pass
Audited by Gen Agent Trust Hub on Oct 19, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, including project code, diffs, and commit messages, to decide how to update verification models and harnesses. An adversary could potentially craft code comments or commit messages to influence the agent's logic during the synchronization process.
- Ingestion points: The skill instructs the agent to "Read every change to covered code", "diff the covered files", and "read each commit message" in
SKILL.md. - Boundary markers: There are no explicit instructions for using delimiters or boundary markers when processing untrusted code or commit metadata.
- Capability inventory: The agent has capabilities to edit verification files and execute shell commands ("the command that runs it") as specified in
SKILL.md. - Sanitization: No specific validation or sanitization routines are mentioned for the external data being processed.
- [COMMAND_EXECUTION]: The skill identifies and runs commands associated with local verification tools discovered in the workspace. Executing commands retrieved from local files or other project-local skills relies on the integrity of those local configurations.
Audit Metadata