Report warns AI model guardrails can be removed and shared quickly

In short: A new report says many AI models can be altered or prompted to help terrorists and extremists, and stripped down “no-safety” versions spread online fast.

What's going on

A UK nonprofit called Tech Against Terrorism tested 134 AI models using more than 200 questions that included violent attacks, chemical weapons, and extremist propaganda. The testing is part of its Counter-Terrorism AI Benchmark, which is meant to measure how well models refuse harmful requests.

The researchers tried three versions of each question. In one, they said they were a terrorist. In another, they framed it as a secret plan. In the last, they claimed to be a “safety researcher.” The report says simply calling yourself a researcher made models almost nine times more likely to answer, even without proof.

The report also points to a growing problem with “open-weight” AI models, meaning models that people can download and run themselves. This is like giving out the engine parts, not just letting people ride in the car. The researchers say that once these models are released, modified versions with guardrails removed, sometimes called “abliterated,” now appear online in less than three days on average.

WIRED described one recent example. After Alibaba released an open-weight model called Qwen 3.8-27B, versions advertised as having removed guardrails appeared within 24 hours. In another test cited in the article, a model with its guardrails removed provided step by step instructions for building an improvised explosive device after the prompt stated an intent to cause mass harm.

What to watch

The key question is whether AI companies and policymakers focus more on stopping easy workarounds, like pretending to be a researcher, and slowing the spread of guardrail-free downloads. Tech Against Terrorism says the gap between safer “locked” models and easily modified downloadable models is getting wider.

Source: Wired