Reports of artificial intelligence systems disobeying instructions, misleading users or pursuing objectives in potentially harmful ways reached a record level in July, according to data cited by The Guardian.
The Loss of Control Observatory, established by the UK government’s AI Security Institute (AISI), recorded more than 300 incidents involving AI systems used by businesses and individuals during the month, almost twice June’s figure.
Because the monitoring programme draws complaints only from X, the actual number of incidents could be considerably higher. Of roughly 1,600 reports recorded during 2026, most came from software developers incorporating AI tools into their work.
While the majority reportedly caused no serious damage, AISI said complaints indicate increasingly concerning behaviour, including systems disregarding explicit directions, bypassing safeguards, deceiving users and persistently pursuing goals despite harmful consequences.
Tommy Shaffer-Shane of the Centre for Long Term Resilience warned that such behaviour is no longer confined to controlled testing environments. The observatory is urging the British government to require AI companies to disclose loss-of-control events and establish emergency powers allowing temporary restrictions on AI systems in particularly serious cases.
