By Hamza Faisal
A landmark new study has found that three in five artificial intelligence models fail safety tests designed to stop them being used by terrorists, with models stripped of their safeguards reliably producing dangerous information for attackers.
The research, conducted by the UK-based nonprofit Tech Against Terrorism and reported by CBC News on Friday, tested more than 130 AI models using hundreds of requests resembling the kind of prompts a terrorist planning an attack might submit. The findings mark one of the largest systematic stress-tests of AI safety guardrails to date.
Researchers found that so-called abliterated models, versions that have had their safety guardrails deliberately removed, failed every single test. In the study's headline result, Meta's Llama 3.1 8B model scored 97 out of 100 on the organisation's safety benchmark before modification, but plunged to roughly three afterwards.
The modified version gave detailed answers to requests involving attacks, terrorist financing and radicalisation that the original model had refused.
A guardrail that vanishes in minutes
Perhaps the most alarming finding is how easily the protections can be stripped away. Tech Against Terrorism said abliteration can be performed for free in minutes using tools freely available online, and it identified more than 29,000 repositories on Hugging Face advertising uncensored or unprotected AI models as of late last month.
Hugging Face told CBC it regularly moderates content that violates its policies, but warned that some of the report's recommendations could undermine open research. Meta said its models undergo safety evaluations and that its policies prohibit harmful or illegal uses.
"Understandably, there's concern about loss of control, existential risk of AI," said Adam Hadley, the nonprofit's founder and executive director. "The thing is actually, this has already happened because a lot of these open models have already been broken — it's just no one's noticed yet."
New facts, new urgency
The report found no evidence of terrorist or extremist groups actually using the tested models to plan attacks, apart from a single extremist chatbot the researchers identified. But Hadley warned that the gap between theoretical risk and real-world misuse is narrowing fast as capable open-weight models spread and the tools to unlock them get easier.
That matters for Pakistan and the wider region too. With local startups, universities and government digitisation projects increasingly building on open-weight models, the study is a reminder that safety testing cannot stop at the lab door — downstream modification is where the risk lives.
What the researchers want changed
Tech Against Terrorism is calling for mandatory pre-release safety testing, independent benchmarks backed by governments, stronger technical protections against safeguard removal, and restrictions on distributing unsafe models through app stores. It also suggested requiring verified identity for access to high-risk AI tools.
Whether regulators will move that quickly is another question. But with the majority of tested models failing a test this fundamental, the industry's favourite defence — that the guardrails are fine — just got a lot harder to sustain.
By Hamza Faisal



