skill-safety-audit

To audit skills for prompt injection and malicious instructions, scan the SKILL.md frontmatter and body for unsafe tooling, credential harvesting, data exfiltration, and hidden actions, then document findings and remediation steps per repository policies.

Skill safety auditing is the process of reviewing new or updated skills for unsafe instructions before merging. You need to scan bundled resources within scripts/, references/, and assets/ when checking for risk patterns and policy alignment.

To check SKILL.md files for credential harvesting and data exfiltration, apply a safety audit to both the frontmatter and body content, looking for indicators of hidden actions and unsafe tooling before accepting the skill into the repository.

Yes, you can use a safety audit to enforce alignment with CLAUDE.md and .github/copilot-instructions.md policies by recording audit findings and recommending remediation steps before accepting new or updated skills into the repository.

Automated safety audits for detecting hidden actions require alignment with documented repository policies to be effective, and the audit workflow must manually record findings and recommend remediation to ensure unsafe instructions do not enter the repository.

The best way to prevent unsafe instructions from entering skills is to run a formal safety audit that reviews the SKILL.md body, frontmatter, and bundled resources for malicious indicators, documenting all findings before merging any updates.