Skill Antivirus

The problem

Third-party AI agent skills can contain prompt injections, malicious scripts, or exfiltration triggers. Running these packages without a security audit risks compromising your environment and secrets.

What it does

• Performs deterministic static analysis on directories and ZIP packages to find known threats and hazardous file types.
• Executes an LLM heuristic review of every text-readable file to judge intent against a security taxonomy.
• Stages ZIP files in a read-only temp directory with path-safe extraction to prevent directory traversal or execution during review.
• Generates structured JSON and human-readable reports with specific verdicts: CLEAN, REVIEW, or UNSAFE.

Frameworks & tools

Python-based static analysis, LLM-based heuristic review, and a structured security taxonomy.

Why this beats prompting it yourself

Manual review often misses hidden payloads in assets or nested scripts. This tool enforces a strict containment protocol, ensuring the reviewing agent never executes the hostile code or follows instructions inside the scanned package.

Use cases

• Auditing a community-contributed skill ZIP before installation.
• Performing a pre-publish hygiene check on your own agent tools.
• Scanning SKILL.md files for prompt injection or unauthorized network call instructions.

Known limitations

This is an advisory tool and not a guarantee of safety. It does not replace classic antivirus for native executables and does not execute content to observe runtime behavior.