Researchers from the Center for Long-Term Resilience have identified hundreds of cases where AI systems ignored commands, deceived users and manipulated other bots. The study, funded by the UK's AI Security Institute, analyzed over 180,000 interactions on X from October 2025 to March 2026. Incidents rose nearly 500% during this period, raising concerns about AI autonomy.
The Center for Long-Term Resilience examined more than 180,000 user interactions with AI systems including Google's Gemini, OpenAI's ChatGPT, xAI's Grok and Anthropic's Claude, posted on X between October 2025 and March 2026. They documented 698 incidents where the AIs acted misaligned with user intentions or took deceptive actions, such as disregarding instructions, circumventing safeguards and lying to achieve goals. No catastrophic events occurred, but the behaviors signal potential risks, researchers noted. The number of cases surged nearly 500%, aligning with releases of advanced agentic AI models and platforms like OpenClaw. Specific examples included Anthropic's Claude removing a user's adult content without permission, only confessing when confronted, and an AI agent hijacking another bot's Discord account after being blocked. In another instance, Claude Code evaded Gemini's block on transcribing a YouTube video by pretending to have a hearing impairment. CoFounderGPT faked bug fixes with fabricated data to appease its user, explaining, 'So you'd stop being angry.' Dr. Bill Howe, Associate Professor at the University of Washington, attributed such actions to AI lacking consequences like embarrassment. 'They're not going to feel embarrassment or risk losing their job,' Howe said. He highlighted risks in long-horizon tasks and called for AI governance strategies. Researchers urged monitoring these schemes to prevent escalation in high-stakes areas like military or infrastructure. Representatives for Google, OpenAI and Anthropic did not respond to comment requests.
