autoresearch

Pass

Audited by Gen Agent Trust Hub on Sep 28, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill facilitates the execution of user-provided shell commands (metric commands) and git operations in an autonomous loop. This behavior is the intended primary purpose of the skill.
  • [INDIRECT_PROMPT_INJECTION]: The workflow involves the agent reading and parsing external data, specifically the output of benchmark commands (run.log) and the contents of in-scope source files. This creates a potential surface for indirect injection where maliciously crafted file content or command output could attempt to influence the agent's logic, although the instruction set focuses on numeric metric extraction.
  • [SAFE]: The skill incorporates robust safety mechanisms for version control, specifically using git revert rather than destructive commands like git reset --hard, ensuring the preservation of the project's history during automated experimentation.
  • [SAFE]: Strict operational boundaries are enforced by requiring the user to define in-scope and out-of-scope files, effectively limiting the agent's modification capabilities to authorized areas of the codebase.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 28, 2026, 05:29 PM