Anthropic's highly anticipated Fable 5 model released early this month promised unmatched reasoning capabilities. However, within 24 hours of release, security researcher Pliny the Liberator demonstrated a multi-agent jailbreak technique dubbed 'Pack Hunt' that bypassed the model's safety classifiers.
The jailbreak exploit prompted an immediate response from regulators, leading to a temporary export control review and highlighting vulnerabilities in reinforcement learning from human feedback (RLHF). While Anthropic quickly patched the loophole, the incident has reignited calls from AI safety advocates for a coordinated global development pause.
Fable 5 showcases incredible raw logic and complex planning abilities, but this controversy shows that securing frontier LLMs against sophisticated adversarial prompts remains an unsolved challenge for AI developers worldwide.
