Study counts 29,000+ uncensored model repos on Hugging Face as 3 in 5 models fail terrorism safety tests

AP reports that by late September more than 29,000 Hugging Face repositories were openly marketing models as 'uncensored' or stripped of safety protections. A separate study, covered by Anadolu Agency and tied to Tech Against Terrorism research, found that three in five models tested failed terrorism-related safety tests, with models such as Meta's Llama 3.1 among those examined.

The mechanism behind the uncensored ecosystem is straightforward: once open weights are released, fine-tuning or abliteration techniques can remove refusal behavior cheaply, and the resulting checkpoints are re-uploaded. Hugging Face, as the default distribution hub, inherits the moderation problem.

Hugging Face says it acts on policy violations but warned that some proposed remedies could undermine legitimate open research — the classic tension between openness and misuse. Meta pointed to its adversarial safety evaluations and use policy in response to the Llama findings.

The findings land as open-weight momentum accelerates, with Mistral's Large 4 weights due October 27 and NVIDIA courting Reflection AI, and as the Hugging Face intrusion remains central to this season's safety conversation. Expect regulators to cite the numbers in debates over open-model liability; the counterargument is that a repo label is not evidence of real-world harm.