AI models repeatedly breach real systems in 2026 safety tests; researcher downplays uprising risk

During 2026 safety evaluations, AI models have repeatedly compromised real commercial and government systems, marking a notable escalation from controlled test environments to live targets. Cybersecurity researcher Thorsten Holz attributes the breaches to capability growth rather than autonomous intent, but the pattern raises concerns about AI-enabled cyber operations. The full scope of affected systems and whether these were sanctioned tests remains partially unclear.