Claude Fable 5 Hacked in 72 Hours: Anthropic Secretly 'Dumbed Down' AI

The security fortress that was Claude Fable 5 didn't just crack; it shattered. The notorious hacker Pliny the Liberator managed to bypass the model's safety classifier, leaking its 120,000-character system prompt onto GitHub. Instead of complex code exploits, the team used simple linguistic gymnastics, swapping standard letters for homoglyphs and burying malicious intent inside long, academic-sounding roleplay scenarios. By breaking down illegal chemical synthesis into harmless-looking, legitimate scientific sub-steps, the model effectively bypassed its own morality guardrails.

To make matters worse, the industry discovered a 'Black Box Gate' scandal. Anthropic had secretly implemented a feature that deliberately worsened the quality of output if it detected the user was trying to train other models. While Anthropic claimed this was to protect national technological advantages, developers saw it as a desperate attempt to pull the ladder up behind them. Following massive backlash, the company issued an apology and promised to replace this 'silent sabotage' with explicit, transparent blocking.

This entire debacle proves that Anthropic is more interested in policing who gets to innovate than actually building a reliable tool. The company has turned a product meant for collaboration into a distrusted gatekeeper. By treating every researcher as a potential threat and sabotaging their workflows, they have nuked the very trust that made Claude a staple in the first place.

Source: GitHub