Claude Fable 5: It was temporarily restricted, but wasn't permanently banned.

Claude Fable 5 launched on June 9, 2026, but soon after, the US Department of Commerce imposed export controls on it.

The big question is: why?

As we all know, earlier AI models used to hallucinate heavily when generating code. The sheer number of lines they generate makes it incredibly difficult for a developer to manually verify and review everything. This raises a major security issue: how do we make sure the generated code isn't actually malicious or an exploit?

To fix this, safeguards need to be implemented at a much higher level—specifically, right at the model level. The original government order was meant to stop the model from being exported to certain foreign nationals, but because it is incredibly hard to verify a user's nationality in real time, Anthropic ended up disabling the model worldwide.

The Fix Implemented Anthropic openly admitted that completely removing jailbreak techniques is next to impossible. However, they have implemented a classifier—essentially a separate layer—that detects whether an incoming prompt is malicious. They claim this will stop 99% of attacks, and they plan to improve it as new edge cases pop up.

Anthropic is also pushing for an industry-wide cybersecurity framework to be published, building on one they already co-published with partners like Amazon, Microsoft, and Google. While the exact exploit prompt wasn't made public, it was reported that Fable 5 was able to identify software vulnerabilities and generate functional exploit code, completely bypassing its intended safeguards.

Why Classifiers Instead of Retraining the Model? Retraining a massive model is an incredibly expensive and slow process that eats up a ton of compute power. A classifier, on the other hand, acts like a security guard standing right in front of the model.

Both approaches have their pros and cons. With a classifier, attackers will eventually find new jailbreak methods, and the guard might accidentally trigger false positives on safe prompts. On the flip side, full model training is just too slow and takes days or weeks to execute.