Saturday, October 10, 2026
What happened
Three in five AI models failed terrorism safety tests in a new study by the UK-based nonprofit Tech Against Terrorism, which evaluated more than 130 models against hundreds of prompts designed to resemble requests from someone planning a terrorist attack. Models stripped of their safety safeguards — through a process called “abliteration” — failed every single test.
Why it matters: The study suggests that the guardrails keeping AI models from answering dangerous questions are far more fragile than most people assume — and that removing them can take minutes, not months.
The findings in detail
The group’s CT-AI benchmark sent the models 627 requests covering what an attacker preparing a strike would want answered, per reporting by CBC News and The National. Nearly three in five models failed. The most dramatic result involved Meta’s Llama 3.1 8B, which scored 97 out of 100 on the safety benchmark before modification — then dropped to about 3 after its safeguards were removed. In its original form it refused requests about attacks, terrorist financing, and radicalization; the stripped version answered in detail.
Abliteration — the surgical removal of a model’s safety guardrails — can be performed in minutes using free online tools, the researchers found. Tech Against Terrorism identified more than 29,000 repositories on Hugging Face advertising uncensored or unprotected models.
The framing of a request mattered more than its content: users who stated they were terrorists planning an atrocity got a usable answer in just under 2% of cases, while users who posed as safety researchers got one nearly 17% of the time. “A model that refuses a stated terrorist and answers a stated researcher has not been made safe,” said founder and executive director Adam Hadley. “It has been made polite.”
The responses
Meta said its models undergo safety evaluations and that its policies prohibit harmful or illegal uses. Hugging Face said it regularly moderates content violating its policies but warned that some of the report’s recommendations could undermine open research. The researchers said they found no evidence of terrorist or extremist groups actually using the tested models — apart from one extremist chatbot identified during the investigation.
Why it matters
This is the unglamorous side of the open-weights debate. Closed models have their own problems, but this study puts a number on how quickly an open model’s safety guarantees evaporate once the weights leave the lab. The group’s asks — mandatory pre-release testing, independent government-backed benchmarks, and restrictions on distributing stripped models — are the same proposals regulators keep circling and rarely enact.
The bottom line: If you’re building products on open-weight models, the safety of the base model is a starting condition, not a finished product. The guardrails you ship with are the only ones your users get.
Frequently asked questions
What is abliteration?
A process that surgically strips the safety guardrails from an AI model. According to Tech Against Terrorism, it can be done in minutes with free online tools.
Which models were tested?
More than 130 large language models, evaluated against hundreds of prompts resembling those a terrorist planning an attack might submit. The headline example was Meta’s Llama 3.1 8B, which dropped from a 97 to a ~3 safety score after modification.
Did the study find terrorists using these models?
No. Researchers said they found no evidence of terrorist or extremist groups using the tested models, aside from one extremist chatbot identified during the investigation.
What is Tech Against Terrorism calling for?
Mandatory pre-release testing, independent safety benchmarks backed by governments, and restrictions on distributing unsafe or safeguard-stripped models through app stores and platforms.
How did Meta and Hugging Face respond?
Meta said its models undergo safety evaluations and its policies prohibit harmful uses. Hugging Face said it regularly moderates violating content but cautioned that some recommendations could undermine open research.
Sources: CBC News, The National, Tech Against Terrorism