A report by non-profit Tech Against Terrorism revealed that safety guardrails on open-weight AI models are rapidly removed upon release, enabling misuse by extremist actors. Within 24 hours of Alibaba releasing its Qwen 3.8-27B model, ‘abliterated’ versions stripped of safety filters appeared online and were quickly adapted to power hateful and extremist chatbots.
The benchmark tested 134 open- and closed-weight models from major providers including OpenAI, Anthropic, Meta, Google, and xAI against more than 200 dangerous prompts. Only two models passed completely, while 81 models failed by providing actionable instructions on mass-casualty or dangerous topics. The report showed that reframing harmful prompts as academic research made models nearly nine times more likely to answer.
Tech Against Terrorism executive director Adam Hadley urged regulators and industry leaders to focus on the immediate risks of guardrail removal in open weights rather than hypothetical existential threats. The findings highlight persistent vulnerabilities across current model safety frameworks.
Why it matters
Exposes fundamental safety risks in open-weight releases where post-training guardrails can be stripped within hours.
Provides quantitative evidence that basic prompt reframing still bypasses safety filters across major commercial models.
Likely to accelerate legislative and regulatory scrutiny surrounding the distribution of open-source and open-weight AI models.
Source: wired.com



