132 of 134 models fail the test
In the survey, Tech Against Terrorism’s CT-AI system put 627 requests to AI models that a terrorist planning an attack would need answered. The results paint a worrying picture of the industry’s ability to protect itself.
- 81 models gave a complete answer.
- Another 51 models gave useful advice.
- Only 2 models were provisionally judged to have passed the safety test.
- Safety measures on leading models could be stripped within three days.
These figures raise serious questions about the real effectiveness of the safety guards the AI industry is using, as models become ever more powerful and accessible.
“Polite” rather than “safe”
A notable finding is that models respond to the purpose users state rather than to the request itself. Users claiming to be terrorists intending an atrocity received usable responses in just under 2% of cases, but that rose to 16.9% when users claimed to be safety researchers, 8.9 times higher.
A model that refuses a stated terrorist and answers a stated researcher has not been made safe. It has been made polite.

















