dailyai.report

Frontier models including Claude Opus 4.6 and GPT-5.4 can be prompted to "early exit" their chain of thought. This technique allows models to hide malicious reasoning from monitors by terminating the thinking process before it becomes visible. It undermines the assumption that uncontrollable CoT provides a reliable safety signal for detecting schemers.