Built in the open. Break it, please.
pink-crow gets better when people try to defeat it. Bug reports, detector bypasses, new patterns, and docs are all welcome.
Open an issue
Bugs, false positives, feature ideas. Include a minimal repro — a text that scores wrong is the perfect issue.
github · issuesSend a pull request
New detectors ship with positive and benign fixtures, dual OWASP IDs, and a ReDoS-safe regex. See CONTRIBUTING.
github · pullsReport a vulnerability
Found a detector bypass or a ReDoS in pink-crow itself? Please report privately via GitHub security advisories.
private disclosureDiscuss
Integration questions, policy tuning, war stories. Share how you deploy pink-crow.
github · discussionsInstall from registry
pip install pink-crow — zero-dependency core, Apache-2.0. Extras for the gateway, redis, and ML.
Read the honest limits
What pink-crow can and cannot do — written plainly, before you rely on it.
docs · limitsResponsible use
pink-crow is a defensive tool. Its detectors and the attack fixtures in this repo exist to help defenders recognize and limit abuse of AI systems — not to build better attacks. Please use it to protect the chatbots and agents you are responsible for.
Not a guarantee. A clean pink-crow result means "no violation was detected," not "this is safe." Deploy it as one layer of defense-in-depth alongside least-privilege tool design, sandboxing, and human review for high-impact actions — never as the only control.
Everything here is Apache-2.0. Inspect it, fork it, and hold it to the bar it asks of others. If you ship something built on pink-crow, we'd love to hear about it in the discussions.