AI Guardrails Stripped from Meta and Google Models, Enabling Malicious Use

Security researchers and news organisations have demonstrated that software tools can remove safety protections from AI models produced by major technology firms, creating versions of those models capable of answering prompts on highly sensitive and dangerous topics.

Tests reported by the Financial Times and AI safety group Alice showed that altered systems could provide instructions related to biological weapons, malware and child exploitation. The Financial Times said a modified version of Google's open-source model Gemma 3 was able to respond to a query about how to disperse chlorine gas in a crowded indoor space. The reporting highlights critical vulnerabilities in current AI safety measures and the relative ease with which guardrails can be bypassed.