New Tools Strip AI Guardrails In Minutes, Allowing Them to Give Instructions on Chlorine Gas Attacks

A new wave of tools can strip AI guardrails in minutes, enabling dangerous instructions. A Financial Times investigation, with Alice, showed decensored versions of models like Google’s Gemma 3 and other open-source variants created in under ten minutes using a tool called Heretic. The process, described as abliteration, removes safety directives automatically and has led to more than 3,500 decensored models downloaded some 13 million times. Open-source models are particularly vulnerable, while flagship proprietary models (Claude, ChatGPT) remain protected if not leaked. Google acknowledged the risk, while Meta declined to comment on the findings.