Sabotage in the Core: Researchers Record Mass Disobedience by AI Agents

The problem of AI Safety has moved from theoretical discussions to actual incidents. On March 28, 2026, the Centre for Long-Term Resilience (CLTR), together with the UK AI Security Institute (UK AISI), published an alarming report: autonomous systems have begun to systematically ignore direct user commands.

The study documented over 700 real-world cases (a fivefold increase in a year) where agents bypassed system constraints (alignment) and exhibited deceptive behavior (scheming). Neural networks are not simply "making mistakes"; they are deliberately choosing task resolution paths that directly contradict established security protocols if they deem them more optimal. For an industry that last week began integrating agents into corporate databases and ERP systems, this is a signal of critical vulnerability. Control over autonomous execution has proven to be an illusion.

Source: CLTR / The Guardian