What this is and how it differs from a manual audit.

What it is

skillsec.io is a security scanner for AI-agent skill files (commonly named SKILL.md). It runs an open, heuristic ruleset against the content of a skill and returns a score plus a list of findings. The whole thing runs in-memory, per request - there is no database, no signup, no API key, and no payment.

Free

No account, no quota, no upsell. Paste, scan, share.

Open ruleset

Every detection is in the source tree. Audit our audits.

Instant

No queue, no callback. One HTTP round-trip per scan.

Coverage

The scanner is content-agnostic. It reads any plaintext file an AI coding agent might obey, including:

  • Claude Code SKILL.md files
  • Cursor .cursorrules and .cursor/rules/*.md
  • GitHub Copilot .github/copilot-instructions.md and .github/instructions/*.md
  • OpenAI Codex / GPT custom instructions
  • Gemini CLI GEMINI.md and prompts
  • Aider CONVENTIONS.md
  • Continue.dev configs and rules
  • Any other plaintext system prompt or agent instruction

The detection rules check for things like curl | sh, secret-shaped env var reads, and prompt-injection phrases. None of them care which agent the file was written for, the threats are the same across the board.

How it differs from a manual audit

Manual security audits are useful and we can't replace them. They're also expensive, slow, and opaque - a human-priced review by a security firm doesn't scale to the number of skills now being published every week. skillsec is the opposite trade-off:

Manual auditskillsec.io
Turnarounddays–weeksseconds
Cost$$$free
Reproducibledepends on reviewerdeterministic
Transparentreport, not methodevery rule is open
Depthdeep, contextualheuristic, surface-level
Coverage at scalebounded by humansbounded by compute

Use skillsec to triage and gatekeep. Use a manual audit for anything you'd trust with elevated privileges or a real user's data.

Open by default

We publish the threat model - every category we evaluate, the severity it can produce, and the scoring weights - on the methodology page. We do not publish the exact patterns: that would hand attackers the cheat sheet for evading the scanner. The source lives in the GitHub repository; if you think a category is missing or wrong, file an issue.

What we don't do

  • We don't store the content you scan. We don't log it. We don't train on it.
  • We don't require an account. There is no account.
  • We don't guarantee correctness. Heuristic detectors are wrong sometimes - both false-positive and false-negative.
  • We don't replace a human security review. Treat a clean skillsec score as a smoke test, not a clearance.