coding-agent-supervision
Pass
Audited by Gen Agent Trust Hub on Sep 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to supervise autonomous agents like Grok Build and Antigravity, which involves ingesting logs and session updates. This creates a surface for indirect prompt injection from the output of the supervised agents. However, the skill provides extensive guidance on verification, boundary markers, and isolation to mitigate this risk. Ingestion points include session logs and update files. Capabilities include file-system operations and testing tools. Boundary markers include isolation worktrees and explicit side-effect prohibitions. Sanitization is recommended for sensitive runtime state and API keys.
- [COMMAND_EXECUTION]: The skill describes a defensive command-wrapping technique to restrict sub-agent capabilities. In references/agy-print-run-audit-2026-09-03.md, it suggests prepending a safe bin directory to PATH to intercept git calls and block dangerous subcommands while allowing safe ones. This environment manipulation is documented as an isolation measure for supervising potentially untrusted agent behavior.
Audit Metadata