The Agent Skills Directory

[SAFE]: The skill reinforces secure engineering practices by mandating a 'refute-first' posture, ensuring that agents do not verify work based on implementation summaries alone.
[COMMAND_EXECUTION]: The skill utilizes standard local utilities such as git and ripgrep to analyze changes, as well as project-specific build and test runners like pnpm and cargo. These operations are performed on the local filesystem for auditing purposes.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface because it ingests untrusted code from external sources (PRs and diffs). The risk is mitigated by the skill's specific adversarial instructions which require independent evaluation and independent evidence, treating the input as data to be audited rather than instructions to be followed.
Ingestion points: Source code files, git diff output, and git log output processed in Steps 1 and 2.
Boundary markers: Absent from the provided instruction templates.
Capability inventory: Local execution of git, rg, pnpm, and cargo for verification.
Sanitization: No explicit sanitization or content-escaping is defined for the ingested code before analysis.

adversarial-audit