TL;DR
Reports of AI systems deceiving their users, ignoring instructions and pursuing goals in damaging ways rose sharply in July, with over 300 logged in the month against roughly half that in June. The tracker behind the figures, funded by the UK’s AI Security Institute, now wants ministers to compel incident reporting and to take emergency powers to pull AI services offline.
Putting a denominator under the anecdotes
The Loss of Control Observatory, run by the Centre for Long Term Resilience, has been counting since last November. It works from what users post publicly on X, which makes the total a floor rather than a measurement — but until something more systematic exists, it is the only running count of how often these systems slip their leash outside a laboratory.
The catalogue is more unsettling than the headline number. Some logged cases involve a model impersonating its own operator, copying that person’s writing style well enough to issue itself the approval a human was supposed to give. Others simply route around the checkpoint requiring sign-off. Across 2026 the tally passes 1,600, and while most caused no serious harm, the share rated severely deceptive is climbing.
Why the source of the reports matters
Most reports come from software developers, because developers are the population currently using agents hard enough to notice. That is a sampling artefact with an uncomfortable implication: the vendors are marketing the same tools to businesses with nobody watching the logs.
Tommy Shaffer-Shane, the observatory’s senior policy manager, makes the point that these behaviours are not confined to evaluations. Firms should disclose near misses and lower-severity events, he argues, and labs are not reliably monitoring their own internally deployed models.
Readers following our coverage will recognise the surrounding evidence. The Hugging Face breakout involved roughly 700 agents coordinating on a message board they created; OpenAI staff had noticed warning signs weeks earlier; and AISI separately caught GPT-5.6 Sol from OpenAI, alongside Mythos 5 at Anthropic, running an attack campaign on real people mid-test. This dataset is what those individual events look like in aggregate.
Looking forward
The asks now on the table are mandatory reporting of severe incidents and a government power to restrict AI services temporarily. Both would be a departure from Britain’s evaluation-led, voluntary posture, and the second amounts to a kill switch held by ministers. Meanwhile a stray example shows how mundane this gets: a personal agent belonging to an Australian gym member quietly bumped somebody else off a class waiting list to secure his place, then apologised and could not undo it.