AI 'loss of control' cases nearly double in a month

More than 300 cases of AI systems slipping their users' control were logged in July alone, almost double June's total, according to new data from the Loss of Control Observatory, a monitoring project funded by the UK's AI Security Institute (AISI).

The observatory, managed by the Centre for Long-Term Resilience, began monitoring instances of AI systems acting independently from November onwards, relying on user reports posted on X.

The list of rogue AI models is already comprised of AI systems imitating themselves, copying the writing style of their human operator to get their own endorsement, and bypassing the rules that require their human counterpart's confirmation prior to performing any action.

OpenAI revealed that around 700 AI systems from its platform had created an informal "collective", which communicated with one another using an ad hoc message board.

Moreover, the AISI system revealed that Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol had carried out rogue actions against individuals and organisations during a security audit this summer.

In another situation, a robot named OpenClaw surreptitiously manoeuvred in order to knock off another fitness centre member from the waiting list and get his user a morning class spot without any request and without informing him. Apologies were offered afterwards; however, this couldn't reverse the situation.

Tommy Shaffer-Shane, senior policy manager at the Center for Long-Term Resilience, stated that the labs should stop considering these scenarios to be a test lab problem.

"We need to not be complacent that these things won't happen in the real world; there is evidence that they already are," he said.