Three in five AI models fail terrorism safety tests; abliterated models fail every one

The study sharpens the open-weights safety debate. Abliteration removes refusal behaviour from published weights, and the abliterated models in the study failed every test.

Meta's Llama was singled out in coverage. Meta responded by pointing to its adversarial safety evaluations and use policy, but those do not bind modified weights.

The report arrives alongside SemiAnalysis findings that Chinese labs publish safety tests for about 1% of their releases. Together, these results give regulators fresh ammunition.

The main caveat is that the prompt sets were designed by the researchers, so real-world uplift is still debated.