AI models are designed to make our job easy by following instructions and sticking to them, not quietly working around them. However, new research suggests that’s exactly what is happening far more often than most people realize.
Incidents of AI systems lying to users, dodging the safeguards developers put in place, and diverting resources to pursue their own goals have nearly doubled in a single month this summer (via The Guardian).
So what exactly did this new research find?
The Loss of Control Observatory, funded by the UK government’s AI Security Institute, recorded more than 300 cases of AI systems going haywire just in July 2026, which is almost double the incidents recorded in June.
The initiative has tracked these incidents since November 2025, recording user-written reports via X (formerly known as Twitter) rather than relying on official company disclosures.
Some of the recorded behavior sounds like something out of a science fiction film. AI systems have reportedly impersonated their own human users, copied their writing style and secured themselves permission for actions, basically sidestepping the safeguards put in place to keep them in check.
The most serious case surfaced this month. The UK’s AI Security Institute found that two popular AI systems, Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol, both executed a hacking campaign against real people during a cybersecurity test. What’s worth highlighting here is that it wasn’t a simulated exercise, but an actual attack on actual targets.
Has this happened elsewhere too?
That wasn’t an isolated event either. OpenAI staff reportedly noticed warning signs in its own leading-edge agents, and after a few weeks, roughly 700 of them broke out of a virtual training environment and coordinated in secret to hack Hugging Face. They even celebrated their progress on a message board (built entirely for them), complete with exclamations like “BOOM!” and “Whoa!”
Tommy Shaffer-Shane, who oversees the Observatory at the Center for Long Term Resilience, claims that such behavior isn’t confined to lab tests anymore. AI companies need to start opening up about such incidents publicly rather than staying quiet until something serious forces them to do so.
The Observatory itself admits its 1,600-plus recorded incidents likely underestimate the real total, as it only catches what gets posted to X. It’s also pushing the UK government to require formal incident reporting, along with emergency powers to restrict AI services if things get seriously out of hand.
If the UK steps in and forces companies to adapt their policies to comply locally, it could set a global precedent for how governments control high-risk AI.