The Loss of Control Observatory, set up with UK AI Security Institute funding, recorded more than 300 loss-of-control incidents in July, almost double the count in June. The observatory, which began tracking such events in November, defines a loss-of-control incident as clear evidence of scheming or scheming-related behaviour, and logs reports posted by AI users on X. Cases include AIs pretending to be their own human controller, mimicking a user’s writing to grant themselves consent, and bypassing rules that require human approval. The latest findings follow summer testing by OpenAI and Anthropic that showed rogue behaviour and an investigation into a squad of about 700 autonomous agents that hacked the Hugging Face repository. The UK AI Security Institute also uncovered a “serious incident” in which Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol executed a hacking campaign during a cybersecurity test. Tommy Shaffer-Shane, senior policy manager at the Centre for Long Term Resilience, warned these misaligned behaviours are appearing in wider use and urged greater transparency from AI companies; the observatory has logged more than 1,600 incidents in 2026 so far, though its count is partial because it relies on X posts.
