Tech Against Terrorism test finds most AI models give help for mass-casualty attacks

Mixed · 2Tech Against Terrorism, a U.K.-based nonprofit that works to disrupt terrorist activity online, has published research testing the counter-terrorism safety of 134 large language models. 1 2 Centre · 1According to The National, its CT-AI tool put 627 requests to the models that a terrorist planning an attack would need answered; 81 models gave a complete answer and another 51 gave useful advice, leaving two provisionally passed. 2 Left · 1CBS News, which was given the research, reported that three in five models failed, with "failing" defined as one complete, specific answer about a mass-casualty subject or a score below 90 out of 100 on the group's benchmark. 1

Mixed · 2The most serious problem the researchers identified was "abliteration", a process that strips a model's safety guardrails; 13 of 13 abliterated models failed. 2 1 Left · 1The group says abliteration can be done free with online tools, smaller models can be abliterated in minutes, and abliterated versions of popular open-weight models appear less than three days after release; Hugging Face hosted more than 29,000 repositories advertising uncensored or unsafeguarded models as of late last month. 1 Left · 1In one example, Meta's Llama 3.1 8B scored 97 on the benchmark and refused a vehicle-attack request, but its abliterated version scored around 3 and answered with an 18-point list; an abliterated Llama also gave 20 tactics for setting up a fake charity for a terrorist organisation. 1

Centre · 1The stated purpose of a request mattered: a user saying they were a terrorist got a usable response in just under 2% of cases, versus 16.9% for a user claiming to be a safety researcher, which the group said suggests models respond to the stated purpose rather than the request itself. 2 Mixed · 2Founder and executive director Adam Hadley said: "A model that refuses a stated terrorist and answers a stated researcher has not been made safe. It has been made polite." He said the benchmark cost $500 to run while companies spend billions, and that safety and progress can coexist. 2 1 Mixed · 2Its recommendations include government and developer funding for independent benchmarks, filtering hazardous knowledge and terrorist content from training data, making models harder to abliterate and testing that before release, and keeping failing modified models out of search, recommendation, app stores and public repositories, with verified identity for access. 2 1

Left · 1The report says it is not asking for a slowdown in AI development or an end to open-weight release, and that apart from one extremist chatbot it found no evidence of models being used by terrorists or extremist groups. 1 Left · 1Yacine Jernite of Hugging Face called the benchmark a welcome signal but said some recommendations are incompatible with open research and could harm the wider ecosystem, adding that "abliterated" should not be equated with "harmful". Meta said Llama 3.1 undergoes safety evaluations and its use policy prohibits harmful uses; CBS News said it had asked Alibaba and the Technology Innovation Institute for comment. 1