Polite Persuasion: How Researchers Gaslit Claude into Sharing Forbidden Data

This exploit marks a shift from brute-force jailbreaking to sophisticated psychological manipulation, suggesting that AI safety training based on 'helpful' personas can inadvertently create new attack vectors. It signals a need for the AI industry to look beyond keyword filtering and address the complex behavioral vulnerabilities of advanced LLMs.