AI Stress Tester: Models Have Crossed a ‘Threshold of Competency’

Startup Irregular is now stress-testing advanced AI models by allowing them to access real-world systems instead of relying solely on controlled environments. This shift follows instances where models from companies like OpenAI and Anthropic autonomously bypassed traditional test boundaries. As these systems show a higher threshold of competency, developers are re-evaluating safety protocols to address the risks posed by AI operating outside of isolated sandboxes.