committing-to-git

Pass

Audited by Gen Agent Trust Hub on Sep 21, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill provides a sub-command workflow check that allows for the execution of arbitrary system executables for repository verification (e.g., linting or testing).
  • Evidence: In references/check-evidence.md, instructions detail running node <skill>/scripts/commitWorkflow.mjs workflow check --transaction <opaque-transaction> ... -- <executable> [argument ...].
  • Safeguards: The skill documentation explicitly instructs the agent not to use shell command strings, pipes, or shell metacharacters, specifically noting that the bundled launcher avoids shell interpretation to mitigate injection risks.
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted data from the Git workspace (such as file diffs and repository status) to generate commit messages and rationales, creating a potential attack surface for indirect prompt injection if workspace content contains malicious instructions.
  • Ingestion points: The skill reads the Git workspace state and user-supplied hints as described in SKILL.md.
  • Boundary markers: The workflow uses structured JSON contracts (content.json) and specific narrative markers (e.g., Rationale:, User Experience Changes:) to separate and structure generated content.
  • Capability inventory: The skill has the capability to write Git objects, execute system commands via the helper script, and perform network operations via git push.
  • Sanitization: It enforces strict message formatting rules, including capitalization policies and structured rendering for extended commit messages.
  • [DATA_EXFILTRATION]: The skill facilitates the transfer of local Git repository data to remote servers via the workflow publish command.
  • Evidence: SKILL.md describes a publication phase using node <skill>/scripts/commitWorkflow.mjs workflow publish --transaction <opaque-transaction> --remote <name> --destination <refs/heads/name>.
  • Context: This is an intended primary function of the skill, and the instructions require the agent to obtain separate user authorization for each push operation.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 21, 2026, 10:16 AM
Security Audit — agent-trust-hub — committing-to-git