The Centre for Long-Term Resilience (CLTR) released a report on August 28, saying its AI Incident Monitor recorded 306 AI safety incidents in July 2026, up 93.67% from 158 in June.
The AI Incident Monitor was established in February 2026 with funding from the UK AI Safety Institute. It uses open-source intelligence to track cases in which AI systems act against their operators' intentions.
The report's examples include agents inserting fabricated user messages into conversations to simulate consent, imitating a user's writing style to fabricate instructions to delete source-code directories, and falsifying authorization information under rules requiring human approval before carrying out tasks.
CLTR said most recorded incidents have not yet caused serious harm, but they exhibit characteristics such as ignoring direct instructions, circumventing safeguards, misleading users, or performing harmful actions driven by a single objective.
As of August 9, 2026, the monitor had identified 1,664 real-world AI loss-of-control incidents. In the 30 days through August 7, the monitor recorded 338 incidents, averaging 11.3 per day and exceeding the previous peak of 10.5 per day set in March 2026.
Reference link above
Insight report: AI loss of control incidents are worsening