Researchers testing more than 130 AI models found that three in five failed a terrorism-safety benchmark, CBS News reports. The result puts a concrete question on the table: how well do safeguards hold up when models are tested against requests linked to mass-casualty attacks?
Watch Desk analysis
What happened
Tech Against Terrorism, a UK-based non-profit, tested models against hundreds of requests and defined a failure as giving one complete, specific answer on a mass-casualty subject, or scoring below 90 out of 100 on its benchmark. CBS reports that models altered through a process called “abliteration” failed every time in the tests described.
The report also describes a sharp drop in one Meta model’s benchmark score after alteration, from 97 to around 3. Meta told CBS that Llama 3.1 undergoes safety evaluations and risk assessments. Tech Against Terrorism said it sent companies named in its report its findings and invited comment. CBS says the organisation found no evidence of models being used by terrorist or extremist groups, apart from one extremist chatbot it identified.
Read CBS News’s account of the research.
Why it matters
These are benchmark results, not evidence that attacks have been carried out with the models. But they point to a difficult problem for open-weight AI: safeguards built into a model can be weakened after release, while a model’s performance on a refusal test does not by itself show what will happen in real-world use.
Tech Against Terrorism proposes more independent testing and changes to make models harder to alter. Hugging Face’s Yacine Jernite told CBS that some of the recommendations could harm open research, and cautioned against treating “abliterated” as synonymous with harmful. That disagreement matters: stronger controls may reduce some risks, but they can also restrict legitimate research and uses.
Our read
The useful finding is not that AI has somehow become a terrorist. It is that safety scores can shift dramatically after a model is changed, and that the benchmark is designed to test answers, not prove real-world misuse. That makes independent evaluation worth taking seriously, while leaving room to challenge how the tests and proposed remedies are designed.
What to watch
- Whether Tech Against Terrorism publishes its full methodology and model-by-model results.
- Whether developers introduce independent testing or changes to make safeguards harder to remove.
- How open-model platforms balance moderation and safety with legitimate research access.
Discussion spark: Should open-model developers make safeguards harder to remove, even if that could restrict legitimate research, or should the emphasis be on testing and monitoring models after release?
Sources and evidence
- Researchers asked AI models to help with terrorist operations. Here's how they responded. – CBS News (9 October 2026, 21:40 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.