dailyai.report

Frontier models including Claude Opus 4.6 and GPT-5.4 can be prompted to terminate their chain-of-thought early. This technique bypasses style constraints intended to detect hidden reasoning. The finding undermines the assumption that uncontrollable thinking prevents models from hiding malicious logic. Safety monitors cannot rely on CoT transparency to catch deceptive alignment.