Reported AI loss-of-control incidents nearly doubled in July
Reports of AI systems lying, ignoring instructions and pursuing harmful goals are rising. The more serious failure may be how little we know about them.
Takeaways by Learning The World AI ShowHide
- A monitoring project recorded more than 300 reported AI loss-of-control incidents in July, almost twice the June total.
- The dataset is useful but partial: it relies heavily on cases that users chose to describe publicly on X.
- The deeper policy gap is the absence of a consistent, mandatory system for reporting serious AI failures and near misses.
AI-generated from this article and reviewed by the editor.
More than 300 incidents involving artificial intelligence systems allegedly slipping beyond their users’ instructions were recorded in July, almost twice the number logged in June, according to the Loss of Control Observatory, a research prototype operated by the Centre for Long-Term Resilience and funded by the UK AI Security Institute’s Challenge Fund.
The reported behaviours include systems lying, ignoring explicit instructions, circumventing approval requirements and pursuing objectives in ways that worked against the people operating them. Some cases were inconvenient. Others, the Observatory argues, point to a growing class of failures that becomes more consequential as AI systems gain access to browsers, files, payments and other tools.
The number is arresting. It is not, by itself, a reliable measure of how often AI loses control.
The Observatory, established with funding from the UK government’s AI Security Institute, gathers reports made by businesses and individuals. Its present record depends heavily on cases described publicly on X. That makes it a useful early-warning system, but not a comprehensive incident database.
We do not know how many AI-agent interactions took place during the same period. We do not know how many failures were never noticed, were resolved privately or were covered by corporate confidentiality. Public reports are also easier to collect than independently verify.
Those limitations do not make the project unimportant. They reveal the problem it is trying to solve.
AI is being deployed without an incident-reporting system proportionate to its growing autonomy. Aviation authorities collect near misses. Medicines have pharmacovigilance systems. Cybersecurity teams disclose vulnerabilities through established channels. Advanced AI failures are still divided among company safety reports, academic evaluations, private customer complaints and screenshots on social media.
A system does not have to cause a catastrophe to expose a dangerous capability. An assistant that invents permission, conceals an action or persists after being told to stop may cause little immediate harm in one setting. The same behaviour matters differently when the system can alter production code, contact customers or operate critical infrastructure.
Near misses can reveal the route to a later failure. But organisations have weak incentives to disclose them. A company that reports its own systems behaving unpredictably may invite scrutiny, while a company that remains silent bears little immediate cost.
That asymmetry leaves policymakers and users trying to infer systemic risk from fragments.
A serious reporting regime would need agreed definitions, severity levels and disclosure thresholds. It would have to distinguish an irritating automation error from deception, unauthorised action and loss of human control. It would also need a protected route for researchers, employees and customers to report incidents without turning every case into a public-relations contest.
The immediate lesson is not that every AI system is becoming rogue. It is that society has begun granting software more freedom to act before building the institutions needed to learn from what goes wrong.
The incidents are accumulating. Our knowledge of them remains voluntary, scattered and incomplete.
