US officials want Anthropic to ensure Fable 5’s guardrails cannot be circumvented before the model is rereleased, WIRED reports. Security experts say that level of jailbreak prevention may not be technically possible.
The demand matters because it exposes a gap between policy expectations and current AI safety engineering. Models can be hardened, monitored, and restricted, but no widely deployed system is known to be jailbreak-proof.
The fight could shape how governments regulate frontier models with cyber capabilities, especially if agencies require guarantees that vendors cannot realistically provide.