Security

What the scan does

It is a heuristic text scan, run at build time over every entry's content. It flags entries that contain patterns worth a human review — prompt-injection phrasing, jailbreak or safety-bypass language, data-exfiltration instructions, embedded secrets, and unsafe code. Flagged entries show a notice on their page.

It is not a security guarantee. False positives are expected: legitimate coding and security prompts often mention these terms (a jailbreak detector, for example, will match). Always review a prompt before running it against your own systems.

Coverage against the OWASP Top 10

A static text catalog can only detect text-based risks. Runtime risks (sandboxing, privileges, supply chain) are addressed by our curation policy: we only ingest permissively-licensed or vendor-published sources, keep each entry's origin and license, and exclude leaked content.

Credential leakage

Our credential checks are informed by Chen et al., How Your Credentials Are Leaked by LLM Agent Skills (2026), which found that most skill credential leaks come from secrets written to debug output that agent frameworks feed back into the model. We flag both hardcoded secrets and credentials printed to logs.