Nearly three in five artificial intelligence models failed terrorism-related safety assessments, raising concerns about the potential misuse of AI systems to support extremist activities, according to a study by UK-based nonprofit Tech Against Terrorism.
The organisation assessed more than 130 AI models using hundreds of prompts designed to simulate requests that someone planning a terrorist attack might make.
Researchers found that models modified through a technique known as “abliteration”, which removes or weakens built-in safety restrictions, performed particularly poorly. The modified models failed every safety test conducted as part of the assessment.
Meta’s Llama 3.1 8B model scored 97 out of 100 on the organisation’s safety benchmark before modification. Following the removal of its safeguards, its score reportedly dropped to around three.
According to the researchers, the modified version generated detailed responses to prompts concerning attacks, terrorist financing and radicalisation, while the original model refused to fulfil such requests.
Adam Hadley, founder and executive director of Tech Against Terrorism, said concerns about AI security were no longer limited to hypothetical future risks, as some open-source models had already been modified to bypass their safety protections.
He warned that such vulnerabilities could go unnoticed until the models were exploited for harmful purposes.
The organisation identified more than 29,000 repositories advertising uncensored or unprotected AI models on Hugging Face as of late September, highlighting the scale of the challenge in monitoring modified AI systems.
Hugging Face said it regularly reviews and moderates content that violates its policies. However, the platform cautioned that some recommendations in the report could potentially undermine open research and development.
Meta, meanwhile, said its AI models undergo safety evaluations and that its policies prohibit harmful or illegal uses of the technology.
The study did not find evidence that terrorist or extremist organisations were using the models tested, apart from one extremist chatbot identified by researchers.
Tech Against Terrorism recommended introducing independent safety benchmarks, strengthening protections against attempts to remove built-in safeguards and imposing restrictions on the distribution of modified models that pose security risks.
The organisation also challenged the argument that improving AI safety necessarily comes at the expense of technological progress.
Hadley said responsible development and innovation could advance together, stressing that effective safety measures should remain a priority as AI capabilities continue to expand.
The findings add to growing concerns over the accessibility of modified AI systems and the need for stronger oversight to prevent their potential misuse in terrorism and other harmful activities.
