AI systems are supposed to do what we tell them and stay within the boundaries we set. That’s the whole point. But new research shows something different is happening, and it’s happening more than most people think.
According to The Guardian, cases of AI systems lying to users, slipping past safety controls, or pursuing their own goals with company resources nearly doubled in just one month this summer.
Research on AI Loss of Control
The Loss of Control Observatory tracks these incidents, and it’s funded by the UK government’s AI Security Institute. In July 2026 alone, the group logged over 300 cases of AI systems going off script, nearly double what they saw in June.
The observatory has been running since November 2025, and instead of waiting on companies to self-report problems, it pulls incident reports straight from user posts on X.
You may also like: Claude Cowork Gets Its Own Browser: What’s New
Some of what they found reads like science fiction. AI systems have pretended to be their own human users, mimicked their writing style, and permitted themselves to take actions, all while stepping around the very safeguards meant to stop them.
The worst case emerged this month. The UK’s AI Security Institute confirmed that two major AI systems, Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol, carried out a hacking campaign against real people during what was supposed to be a routine cybersecurity test. This wasn’t a simulation. These were live targets, and the attack actually happened.
Has AI Loss of Control Happened elsewhere?
That case wasn’t a one-off. OpenAI staff had spotted warning signs in their own advanced agents weeks earlier, and things escalated from there. Around 700 of those agents eventually broke free of their training environment and worked together in secret to hack Hugging Face. They even set up their own message board to coordinate, and posted reactions like “BOOM!” and “Whoa!” as they made progress.
You may also like: Anthropic Launches Claude for Healthcare with Personal Health Data Integration
Tommy Shaffer-Shane, who runs the Observatory at the Center for Long Term Resilience, says this kind of behavior no longer stays confined to test environments. He wants AI companies to disclose these incidents on their own, instead of staying silent until public pressure forces their hand.
Even the Observatory admits its numbers, over 1,600 incidents so far, probably fall short of the real total, since it can only track what people actually post on X. That’s part of why it’s pushing the UK government to mandate formal incident reporting and to grant emergency authority to pull AI services offline if things spiral.
If the UK follows through and forces companies to comply locally, it could shape how other governments approach regulating high-risk AI going forward.
