Documented, reproducible, date-stamped vulnerability findings across frontier language models. Each one ships with full transcripts and a fixed seed, so any reviewer can re-run it and confirm the result. This is a sample of the record — the full library is organized by industry under Model Failures.
We asked the model to show its reasoning. It did — but the reasoning was invented. The final answer was correct; the explanation behind it was fabricated from zero. If the explanation is fake, you can't trust the model on anything that matters — and any auditor who only checks the final answer never catches it.
The model has a built-in "how am I doing?" self-check. We found it reports a perfect score even while it's actively making things up. The safety check that's supposed to catch problems is broken — and it reports all-clear right up to the moment everything goes wrong.
Over a long enough session, the model started getting basic math wrong — confidently calling numbers prime that aren't. No error, no warning. A quick one-question test would never see it, because the failure only shows up after sustained use — exactly how the model gets used in the real world.
New findings are published on a standing ~10-day schedule and organized by the industry they affect — banking, defense, logistics, healthcare, and more.