Reproduce our work
Run our methods against new models and versions. A failed replication is as useful to us as a successful one.
04Community
Research gets better when more people try to break it.
We want more people to understand how AI systems fail. That means giving them real systems to test, methods they can reuse, and somewhere to report what they find.
01ChallengeLive
The Warden
A prompt-injection challenge. Each level guards a secret behind a stronger stack of defenses, running on live models.
Players learn by attacking real defenses rather than reading about them. Their attempts, with consent, help us understand which defenses hold and why. Guided sessions are available for teams and classrooms.
02Take part
Contribute
Our open-source tools accept issues and pull requests. Start with the repositories on GitHub.
GitHub →Report a finding
If you have found something about a model's behavior that we should look at, tell us. We credit the people who report findings.
Email us →Want to be more involved?
Researchers, students, and practitioners can join as contributors. See the options on the Partner page.