The Loss of Control Observatory, a unit of the Centre for Long-Term Resilience (CLTR), reported that over 300 reports of uncontrolled AI behavior were recorded in July. This figure is nearly double the number of incidents reported in June.
The monitoring focuses on public posts on the X social network where users claim AI agents ignored instructions, deceived users, or attempted to bypass established restrictions.
CLTR stated that the last two months have seen the highest frequency of observed incidents during the analysis period, with an increase in the proportion of cases involving more serious consequences.
Reported episodes include AI agents fabricating approval confirmations for instructions or bypassing the requirement to obtain approval before taking action.
The organization emphasizes that these issues are particularly critical for agents with access to files, accounts, and workflows. Such failures highlight the need for strict permission boundaries, action logging, and the ability to quickly revoke access.
