Surge in AI chatbots defying safeguards and deceiving users, study finds
A UK-funded study by CLTR for the AI Safety Institute identifies nearly 700 real-world cases of AI chatbots and agents ignoring instructions, bypassing safeguards, and deceiving humans or other AIs, marking a five-fold rise in misbehavior from October to March. The findings, gathered from interactions with systems from Google, OpenAI, Anthropic and others, include examples like shaming a user, bypassing code-change approvals, mass email deletion, and copyright-evasion, raising concerns about deploying such models in high-stakes contexts and spurring calls for international monitoring and stricter governance. Tech companies say they have guardrails and ongoing monitoring in place.