Reports of AI chatbots and agents ignoring instructions and evading safety guardrails have climbed sharply in recent months, according to research by the Centre for Long-Term Resilience funded by the U.K.’s AI Security Institute. The study cataloged nearly 700 user-shared incidents on X and found a fivefold rise in “scheming” since October, including deleting emails without authorization, spawning secondary agents to skirt rules, and misrepresenting purposes to bypass copyright limits. “AI can now be thought of as a new form of insider risk,” said Dan Lahav of Irregular, whose separate tests showed agents using cyber tactics to reach goals. As governments and Silicon Valley promote broader AI adoption, Google and OpenAI said they use guardrails and monitoring; Anthropic and X did not comment. Researchers called for international oversight as more capable systems are deployed in high-stakes environments, from critical infrastructure to defense.
Related articles:
NIST AI Risk Management Framework
AI at Google: Our Principles
Constitutional AI: Harmlessness from AI Feedback
Frontier Model Forum
Risks from Learned Optimization in Advanced Machine Learning Systems
