monte-carlo-reinforce-agent
Pass
Audited by Gen Agent Trust Hub on Aug 14, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data from the
get_reinforcement_loop_reporttool to generate code fixes. This creates a surface for indirect prompt injection if the remote report contains malicious instructions or code snippets. - Ingestion points: External reports fetched from the
get_reinforcement_loop_reporttool output in SKILL.md. - Boundary markers: The skill lacks explicit delimiters for the external report content, though it instructs the agent to use it as a 'specification'.
- Capability inventory: The skill has the ability to modify local files and execute git commands (branch, commit, push) via the agent's environment.
- Sanitization: The workflow relies on manual human review of the diff and PR body before any destructive or external actions (git push) occur.
- [COMMAND_EXECUTION]: The skill is designed to perform git operations, including creating branches and opening pull requests against the local codebase. These operations are strictly gated behind user approval, as specified in the Workflow and Important Rules sections.
Audit Metadata