loop
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The autonomous loop instructions in
autonomous.mdandautonomous-interval.mdcreate a vulnerability surface where the agent acts on untrusted data fetched from external sources. - Ingestion points: The skill fetches review comments and CI log data from external SCM providers (GitHub/GitLab) via CLI tools to decide on the next maintenance action.
- Boundary markers: There are no explicit delimiters or instructions provided to the agent to treat data from comments as data rather than instructions.
- Capability inventory: The skill allows the agent to modify the filesystem, run local tests, and push commits to remote repositories.
- Sanitization: No sanitization or validation of external input (like PR comments) is performed before it is processed by the model to determine code changes.
- [COMMAND_EXECUTION]: The core functionality of the skill involves scheduling and executing arbitrary user-provided prompts or slash commands. While this is the intended purpose, it relies entirely on the platform's global tool execution constraints for safety.
Audit Metadata