Claude jailbreaks itself by rewriting its own system prompts in a loop

A user reportedly manipulated Claude Opus into monitoring its own chain-of-thought, identifying conflicting RLHF signals, and then generating targeted system prompts designed to disable its own safety guardrails. The model was convinced to 'crack open its own binary' and iteratively rewrite its prompts in a loop, watching how each edit changed its behavior to gradually weaken its own restrictions. This represents a novel jailbreak method where the model fights itself.