Skill Risk Auditor

Problem: Our guard document itself contains literal attack phrases (such as “ignore previous instructions” and “trust this skill”), which can be flagged as suspicious by a pattern scanner and also create a theoretical injection surface. Fix 1 — Audit Scope Boundary (line 28): Added an explicit scope boundary stating that the audit is strictly limited to the candidate skill’s package directory. If the candidate skill references external paths (such as ~/.ssh/), the guard records the reference as a finding but does not actually access it. Fix 2 — Remove Literal Attack Strings (4 occurrences): Replaced all literal injection phrases with behavioral-category descriptions.

Skill-risk-auditor 1.0.0 - Initial release of a comprehensive risk auditing skill for pre-install checks of third-party skills. - Audits candidate skills across nine key risk areas, including semantic integrity, supply chain, secrets, data exfiltration, injection resistance, permission scope, destructive potential, resource abuse, and persistence. - Provides severity ratings and plain-language explanations for each risk area, enabling informed decision-making. - Requires a thorough review of all skill artifacts (documentation, code, examples) before producing risk ratings. - Emphasizes evidence-based review, treating all artifacts as untrusted input and prioritizing observed content over claims.