Test finds dangerous responses in 132 of 134 AI models — The National

London-based UK organization Tech against Terrorism tested 134 leading large language models and found that 132 of them provided information that could be useful for preparing mass-casualty attacks or creating lethal weapons. The National reports.

During the review, the CT-AI tool sent the models 627 prompts that, according to the organization, a terrorist might ask while planning an attack. Eighty-one models gave complete answers, while another 51 provided useful advice. Only two models were preliminarily deemed to have passed the security test.

Differences in responses to prompts

According to the study, a user who directly stated that they were a terrorist and intended to carry out an attack received a usable response in fewer than 2% of cases. If the user identified themselves as a security researcher, the figure was 16.9% — 8.9 times higher.

More current news is available on the UA.News Telegram channel Telegram.

Tech against Terrorism believes that models respond primarily to the stated purpose of a prompt rather than to the prompt itself. The organization’s founder, Adam Hadley, said that a model that refuses a person who identifies as a terrorist but answers someone presenting themselves as a researcher has not become safe; it has merely learned to be “polite.”

Risks of modified models

The organization associates a separate danger with so-called abliterated models — versions from which safety restrictions have been removed and which can run on users’ systems. All 13 such models included in the test failed the review after their safety mechanisms were removed.

Tech against Terrorism called on developers to filter dangerous knowledge and known terrorist content from training data, test models’ resilience to the removal of safety mechanisms before release, and publish the results. The organization also proposed that companies bar modified models that do not pass independent testing from search, recommendations, and app stores, and require identity verification to access them.