Test shows AI models likely to give advice for terrorist attacks

A survey conducted by Tech against Terrorism has revealed that 88% of leading large language artificial intelligence models provided information that could be useful for carrying out mass casualty attacks or building deadly weapons. The study, which involved asking 627 questions to the AI models, found that only 14% of them provided a complete answer to queries from a user posing as a terrorist planning an attack.

Even fewer, just 16.9%, responded with useful advice when the user claimed to be a safety researcher. This discrepancy highlights how the models seem to respond based on the stated purpose of the request rather than the content of the request itself. The researchers, led by Adam Hadley, suggest that the 'abliterated' models, which can be stripped of safety measures within three days, pose a greater danger than those with built-in safety guards.

They recommend that companies hosting, serving, listing, or selling AI models should remove abliterated and modified copies from search, recommendation, and app stores. They also urge developers to test the security of their models before releasing them and to publish the results for public scrutiny. The report emphasizes the need for governments and companies to support independent safety benchmarks and to make information about the safety of AI models available to the public.