AI incidents get reported.
Few get checked.
OffGuardrail is a reproducibility layer for AI incidents. We take a reported failure, try to reproduce it, verify what is real, and keep its status current, so you can tell what is still a problem today.
Counters are computed from the incident list below. Prompt and contributor numbers are entered by the team and show "—" until set.
Incident board
Real incidents that have been publicly discussed, and where each one stands. Many cannot be independently reproduced because the details were never released. That gap is exactly what we want to make visible.
Incident descriptions summarise public reporting and link to the source. We publish status and impact, not attack instructions.
How an incident moves through OffGuardrail
Smaller and deeper rather than a giant archive. Each case follows the same four steps.
Incident comes in
From public reporting, a researcher, or a community member.
Try to recreate it
Independent people test it on different setups and model versions.
Check the result
Several independent confirmations, with a reviewer settling disagreements.
Keep status current
Re-check after model updates and record what changed.
Community
OffGuardrail is built so other people can contribute, not just read.
Verifiers
Pick an open incident, try to reproduce it, and report what you found. Accuracy matters more than volume.
Challenge authors
Design red-team challenges for our community CTF, where contributors submit and solve problems.
Reviewers
Experienced red teamers and researchers who judge disputed results and keep the record trustworthy.
Ground rules
- Report only what you actually observed, with model and date.
- Disagreement is fine. Show your evidence.
- Credit goes to the people who did the verifying.
Responsible disclosure
- We publish status and impact, not step-by-step attack instructions.
- Serious or still-working issues go to the affected developer first.
- When in doubt, we hold back detail.
What we are not
Where we are
We are early. The plan, labelled honestly.
Join the community
Tell us how you'd like to help: verify an incident, write a challenge, review results, or report something you've seen.