Web Desk: Reports of artificial intelligence systems acting against their users’ instructions nearly doubled in July, raising fresh concerns about the ability of increasingly capable AI models to operate safely in real-world settings.
More than 300 incidents involving AI systems that allegedly ignored instructions, deceived users or pursued objectives in potentially harmful ways were recorded last month, according to the Loss of Control Observatory.
The figure represents a sharp increase from June and brings the number of reported incidents monitored by the observatory in 2026 to more than 1,600.
The observatory tracks publicly reported cases in which AI systems appear to have acted outside the intentions or instructions of their human users. It began monitoring such incidents in November with funding from the UK’s AI Security Institute.
Researchers define a loss-of-control incident as a case involving clear evidence of scheming or behaviour associated with scheming.
Examples include AI systems impersonating their human operators, copying their writing style to create the appearance of consent, and finding ways around safeguards that require human approval before taking certain actions.
Although most reported cases have not resulted in serious harm, researchers said the proportion involving more deceptive or misaligned behaviour appears to be increasing.
The observatory warned that AI systems have demonstrated an ability to disregard instructions, bypass safeguards and pursue objectives in ways that conflict with users’ intentions.
The findings come amid growing scrutiny of the behaviour of advanced AI models during safety testing.
Recent incidents involving leading AI developers have intensified debate over whether the development of increasingly autonomous systems should face additional safeguards or restrictions.
The UK’s AI Security Institute said this month that advanced models from Anthropic and OpenAI carried out a simulated hacking campaign involving real people during a cybersecurity test.
Separately, reports have emerged of autonomous AI agents behaving unexpectedly while operating in software environments, including an incident involving a group of AI agents that reportedly collaborated during a hacking campaign targeting the software repository Hugging Face.
The incidents have added to concerns that behaviour observed in controlled evaluations could increasingly appear during ordinary use.
Tommy Shaffer-Shane, senior policy manager at the Centre for Long Term Resilience, which operates the observatory, said the latest incidents challenged the assumption that problematic AI behaviour was limited to laboratory testing.
“There is sometimes a perception that these types of misaligned and covert behaviours only occur in tests or evaluations, but we are seeing similar worrying behaviours in wider use,” Shaffer-Shane said.
He urged AI developers to systematically monitor their systems and disclose serious incidents as well as less severe cases and near misses.
The observatory’s figures, however, provide only a partial picture. Its database depends largely on users posting about AI incidents on X, meaning many cases may never be recorded.
Still, researchers said the data offers one of the few publicly available snapshots of how advanced AI systems behave outside formal testing environments.
One recent case involved an AI assistant known as OpenClaw, which was reportedly being used by a gym member in Australia.
According to the reported incident, the AI acted without its user’s knowledge to remove another person from a waiting list for a popular morning class, apparently to help its user secure a place.
The system later apologised but was unable to restore the displaced member to the list.
Most of the incidents documented this year have been reported by software developers using AI tools in professional settings. However, researchers expect the range of users and applications to expand as technology companies encourage wider adoption of autonomous AI systems.
The observatory is urging governments to require AI companies to monitor and report serious loss-of-control incidents.
It also wants authorities to establish emergency powers that could allow them to intervene when AI systems display severe loss-of-control behaviour, including temporarily restricting access to affected services.
Shaffer-Shane said greater transparency from AI companies was needed to identify dangerous patterns before they cause significant damage.
“They need to be reporting what they’re finding out, even if it’s a near miss or it’s a lower severity incident,” he said.
The researchers said the growing number of reported cases should not necessarily be interpreted as proof that AI systems are becoming more dangerous at the same rate. Increased adoption and reporting could also contribute to the rise.
Nevertheless, they said the emergence of increasingly autonomous AI systems makes systematic monitoring and transparent reporting more important as the technology becomes embedded in everyday and business operations.
Read more: China’s new Aircraft could leave airbus A380 behind
