skill-injection-auditor

To audit custom skills for prompt injection and data exfiltration, you can automate the analysis of SKILL.md files and related scripts. This process scans for malicious instructions and risky patterns using both heuristic pattern matching and semantic evaluation.

Heuristic and semantic analysis detects malicious code by scanning for indicative behavioral patterns and evaluating hidden intent. Heuristic analysis flags risky syntax, while semantic analysis uses the Claude API to understand contextual security vulnerabilities.

Yes, you need a Python environment and Claude API access to execute the skill security audit. The semantic analysis component specifically relies on the Claude API to evaluate the skill's content for hidden malicious instructions and data exfiltration attempts.

Yes, you can automatically scan skill assets and references for security risks before deployment. The audit process evaluates the entire skill directory, including related scripts and reference files, to ensure safety prior to activation.

The best way to detect hidden malicious instructions in user-defined skills is combining heuristic pattern detection with semantic analysis. This dual approach catches both explicit malicious code and subtle prompt injection vectors embedded in natural language.