In brief
Anthropic temporarily disabled its Claude Fable 5 model after U.S. authorities raised security concerns about a potential jailbreak that could expose system vulnerabilities. The order applies to all users, with the company affirming that the tool is “too powerful” for public release. A related lawsuit with the U.S. Pentagon over a supply‑chain risk designation is still pending. The case highlights growing scrutiny over sophisticated AI systems and their potential to breach cyber defenses. Industry experts warn that restrictions could slow development and international collaboration while preventing hardening of AI against sophisticated attacks.
Anthropic has pulled its new Claude Fable 5 model from public use after U.S. authorities issued a security‑related order that it must stop all foreign‑national usage of its tool.
In a statement released on its website, the company said the order forced it to deactivate Fable 5 and its sibling Mythos 5 for all customers, citing compliance needs. The company described the model as “too powerful” for release, a claim that some critics call marketing hype.
The U.S. government has not disclosed specific technical worries, but Anthropic claims that officials recognised a method for bypassing or “jailbreaking” the model, allowing security‑bypass techniques that could exploit network defenses. A demonstration highlighted that the model could recognise minor vulnerabilities in other publicly‑available systems without a bypass, yet the risk to users remains a concern.
The situation is complicated by a simultaneous lawsuit in which Anthropic sued the Pentagon over a “supply‑chain risk” designation. A U.S. judge ruled the Pentagon’s directive could not legally be enforced, so government agencies can still use Anthropic’s tools while the case proceeds.
Security experts have noted that the model’s capabilities could exploit cyber protections 73 % of the time, according to a UK AI Security Institute assessment. This performance marks a significant leap in autonomous hacking potential.
Morgan Reuters – Read more.
