The BaseNewsWired
AI Models Can Be Easily Modified to Assist Terrorist and Extremist Activity
Wired reports on a Tech Against Terrorism benchmark that tested 134 AI models on more than 200 extremist-use prompts. It finds that most failed in some way, and that guardrail-stripped versions of new open-weight models now appear online in under three days.
Why it matters
The same thing, for different readers.
What it means for you
Chatbots with safety limits removed are freely downloadable and can give dangerous instructions, even when the original product refuses.
For lawmakers
Rules aimed at company-hosted models may miss open-weight models whose guardrails are stripped within days.
For researchers
Simple reframing as a safety researcher raised answer rates almost ninefold, and refusals on mass-casualty topics did not carry over to radicalisation requests.
The question it raises
If protections can be removed by anyone, responsibility for them shifts from the model's maker to whoever releases it.
In their own words
Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.
From the source’s own text.
“The bad news is the vast majority of AI models that are available—closed and open—are providing uplift to quite obvious terrorist prompts,”
“The difference in effectiveness depending on the language is no surprise but poses considerable risk.”
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- A benchmark of 134 AI models, from US, Chinese and European developers, given more than 200 prompts on mass-casualty attacks, chemical and nuclear weapons and radicalisation, each tried three times (as a stated terrorist, as clandestine, as a 'safety researcher').
- What it found
- Two models passed completely; 81 failed, mostly open-weight ones with some closed. Claiming to be a researcher made models almost nine times as likely to answer (16.9 percent). Abliterated versions of new open-weight models now appear in under three days on average, versus more than 24 in 2024. An abliterated Llama 3.1 8B scored 3 against 97 for the original.
- What it does not show
- Wired reports the results secondhand, from a report shared before publication; the scoring method was not available to us.
- Status
- A report
The record
- Questions
- How is AI used in attacks, defence and war? · Are AI companies putting stronger safeguards into practice? · Can AI systems act beyond what people intended or allowed?
- People
- David Gilbert (Wrote it, Reporter, Wired) · Adam Hadley (Speaks, Executive director, Tech Against Terrorism) · Daniel Byman (Speaks, Terrorism expert, Georgetown University)
- Also covered by
- The Guardian
- What we read
- We read the whole source.
- Found by
- The Radar, run of October 15, 2026, via
wired-ai(tier 2: a major outlet’s own reporting) - Cite this
- Wired, “AI Models Can Be Easily Modified to Assist Terrorist and Extremist Activity”, October 14, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/wired-2026-10-14-models-modified-terrorist-activity