skills/matteocervelli/llms/pre-commit/Gen Agent Trust Hub

pre-commit

Pass

Audited by Gen Agent Trust Hub on Sep 2, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the project environment and passes it to secondary AI models for review.\n
  • Ingestion points: Git diffs ($DIFF), staged changes, .forgejo/workflows/ files, and alembic.ini configuration.\n
  • Boundary markers: The skill uses a --- DIFF --- delimiter in Step 6, but lacks explicit instructions for the model to ignore instructions embedded within the code diff itself.\n
  • Capability inventory: The skill utilizes Bash for command execution, Read for file access, and writes to CHANGELOG.md.\n
  • Sanitization: Raw diff content and file data are interpolated into prompts and shell scripts without evidence of escaping or sanitization.\n- [COMMAND_EXECUTION]: The skill supports a --dast <url> parameter which is passed directly to a shell script (dast-runner.sh <url>). This pattern creates a shell command injection surface if the agent does not sanitize the user-provided URL before execution.\n- [PROMPT_INJECTION]: The LIGHTWEIGHT_PROMPT contains instructional overrides such as "Ignore style, naming, structure, and suggestions" and "Focus ONLY on...". While intended to scope the companion review task, these patterns represent the use of instruction-override techniques.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 2, 2026, 11:46 AM
Security Audit — agent-trust-hub — pre-commit