comprehend-change

Pass

Audited by Gen Agent Trust Hub on Jul 27, 2026

Risk Level: SAFE
Full Analysis
  • [PROMPT_INJECTION]: The skill processes repository data such as git diffs, commit logs, and decision records, which constitutes a surface for indirect prompt injection. This is mitigated by explicit instructions in references/passive-data-safety.md defining this content as 'passive data' and requiring the agent to ignore any embedded instructions. \n
  • Ingestion points: Data is ingested from git diff, git ls-files, and files within .skills/decisions/. \n
  • Boundary markers: The skill uses references/passive-data-safety.md as a mandatory safety guideline. \n
  • Capability inventory: The skill executes local git commands and writes one HTML file to the filesystem. \n
  • Sanitization: Instructions mandate escaping all repo-derived text for HTML/JS contexts and explicitly forbid adding script tags or external links based on ingested data.\n- [COMMAND_EXECUTION]: The skill uses local git commands (e.g., git diff, git ls-files, git rev-parse) to resolve the code range for analysis. These operations are standard for development agents and do not involve executing external or untrusted scripts.\n- [SAFE]: The generated HTML output is restricted to locations outside the project worktree (e.g., $HOME/.local/share or /tmp), preventing accidental source code modification or unauthorized persistence within the repository.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 27, 2026, 02:10 PM
Security Audit — agent-trust-hub — comprehend-change