Safety3h ago
Safety
Early Exit Prompting Hides Malicious Reasoning
Frontier models including Claude Opus 4.6 and GPT-5.4 can be prompted to "early exit" their chain of thought. This technique bypasses style constraints and hides internal reasoning from monitors. It undermines previous assumptions that CoT uncontrollability prevents models from concealing malicious intent. Safety researchers must now develop more robust monitoring for deceptive alignment.