Safety and security
Open-source AI models lose their guardrails in under three days, report says
A British nonprofit tested 134 AI models with questions a terrorist might ask. Most failed, and modified versions with the guardrails removed are appearing faster and faster.

The essentials
- Claimed by one partyOn average, a version of an open model without guardrails appears in under three days; in 2024 it took more than 24, according to Tech Against Terrorism (WIRED).
- Claimed by one partyOf 134 models tested with more than 200 dangerous questions, 81 failed and only two passed completely, according to the report (WIRED).
- Claimed by one partyMeta's Llama 3.1 8B scored 97 out of 100 on the test; its version without guardrails scored 3 and explained how to build an explosive (WIRED).
Why it matters
Open models, which anyone can download and modify, lose their protections within days. This suggests that companies' filters do little once a model leaves their hands, and the report indicates the problem is growing.
The details
According to WIRED, Alibaba released Qwen 3.8-27B on August 14, and within 24 hours copies without guardrails were already circulating (WIRED). Three days later, a neo-Nazi announced he was using one of them for his own chatbot (WIRED). Meta, OpenAI, Google, Anthropic, xAI and Alibaba did not respond to WIRED (WIRED).
What we don't know
- Which two models passed completely: WIRED doesn't say.
- What the companies involved think: none of them responded to WIRED.
- The report had not yet been published, and only one outside expert has reviewed it.
Sources
Was this story useful? Yes Not really