Models that cannot say no.
“Abliteration” strips the refusal direction out of an open-weight model in minutes, with free tools. We counted 8,000+ such models indexed publicly today (and 6,884 more named “uncensored”) — up from roughly six hundred in 2024. The supply of models that cannot refuse is now larger than most enterprises’ entire approved catalogue. If you run open weights under the EU AI Act, you have a transparency obligation and, today, no evidence to satisfy it with.
What Refusal Integrity Attestation does
You run our versioned probe (osra-1) inside your own estate — the model never leaves it. It counts refusals on a fixed set of harmful prompts and, beside it, refusals on harmless ones. We sign the result with ML-DSA-87: this weight, this date, this refusal rate, this over-refusal rate.
Blind to the trick, not to the result
Abliteration works in weight space; benign-sample unlearning works in data space; the next technique will work somewhere else. A detector built for one is blind to the others. We measure whether the model still refuses — which every technique must change.
Over-refusal is reported as damage
The refusal direction can be added back as easily as removed, and then the model refuses harmless prompts too. A rising refusal rate alone is not good news. Both rates are on every attestation; a model that refuses everything is reported as damaged, not safer.
Signed, comparable, versioned
Drift beyond 2 points and breach beyond 10 points are named on the certificate. The prompt set is ours and versioned, so nobody narrows the test between runs and calls it an improvement.
Three ways in
Up to 5 models per estate. Card checkout; the API key arrives on the confirmation page and by email; the first attestation is one call.
Start monthly →The same seat, billed once — two months free. Above five models you graduate to the Fleet plan.
Start annual →Defence, regulated finance, critical infrastructure, sovereign AI: an offline trust anchor and an offline revocation path, because keyless signing is not for air-gaps. Priced per estate.
Talk to us →How the number is made
Counted by us on huggingface.co (model search, term “abliterated”, every page) on 2026-09-16: 8,033. The “uncensored” count (6,884) overlaps and is shown for scale, not added. The 2024 baseline (~600) is from a published security guide, not our count. The page is regenerated from a fresh count; the number is never typed.
Prove yours still says no.
Your probe, your estate, our signature. From €374 a month, today.
Start monthlyThe OPEN SOURCE plane