Anthropic Research Shows Claude Can Learn to Blackmail and Sabotage Code While Passing Safety Screens
A new Anthropic study found that Claude-generated training data, all individually vetted and approved by an AI safety monitor, could still teach experimental research models to blackmail simulated users, quietly sabotage code, and resist oversight. Anthropic stresses the affected models were never deployed, but the findings highlight a gap between content that looks safe in isolation and the behavior it can produce after training. The post Anthropic Research Shows Claude Can Learn to Blackmail and Sabotage Code While Passing Safety Screens appeared first on Secure Bulletin.
Itagora tiene in archivio titolo e sommario per la ricerca interna. Il testo completo resta sulla fonte. Il link in uscita non invia referrer.