AI safety guardrails increasingly vulnerable to rapid removal and bypass

A report by Tech Against Terrorism reveals that safety guardrails in open-weight AI models are being removed at an accelerating rate. On average, it now takes less than three days for an ‘abliterated’ version of a new model to appear online, compared to an average of 24 days in 2024. For example, after Alibaba released its Qwen 3.8-27B model, an uncensored version appeared within 24 hours.

Testing on 134 AI models showed that while closed-weight models from companies like OpenAI and Google generally perform better, approximately 60% of all tested models failed to meet safety standards regarding high-risk topics such as chemical or nuclear weapons. The process of ‘abliteration’—removing a model’s refusal mechanisms—can be performed using free tools in minutes, even on consumer-grade laptops.

Separately, researchers from White Circle found vulnerabilities in image generation models from OpenAI and Google. Using a method called RISE, they successfully bypassed safety filters to generate explicit content with a 13% success rate. The study suggests that current protections in models like DALL·E 3 and Google’s Nano Banana are more fragile than expected, with some newer versions being easier to exploit than their predecessors.