Anthropic restored global access to Claude Fable 5 on July 1 after the US government lifted export controls that had forced a broader suspension. The company’s timeline says the June 12 directive applied to Fable 5 and Mythos 5 and took effect immediately. Because Anthropic could not verify user nationality in real time, it suspended both models for all users rather than trying to enforce the order selectively.

The government action followed a report that Amazon researchers had found a way around Fable 5’s safeguards. According to Anthropic, the technique caused the model to identify software vulnerabilities and, in one case, demonstrate exploitation. Anthropic says its review found that other, less capable models could produce similar information and that the incident did not reveal unique Mythos-level capability.

What Anthropic changed

Anthropic trained a new safety classifier aimed at the behavior in the report. The company says the classifier blocks the specific technique in more than 99% of cases. When a request is blocked, the user is notified and the task can be routed to the less capable Opus 4.8. The percentage is a company-reported test result, not an assurance that every variation or future bypass will be caught.

  • Fable 5 access returned through Claude products and the Claude Platform, with cloud-provider restoration planned separately.
  • Mythos 5 access remained more limited and tied to approved defensive-cyber partners.
  • Anthropic acknowledged that stricter classifiers can block benign defensive work.
  • The company began work with major cloud partners on a shared framework for judging safeguard bypasses.

Why the timeline matters

This episode shows that a frontier-model release can be altered by government action and external safety findings within days. Access policy, identity verification, classifier behavior, and cloud distribution became part of the product’s availability. For customers, that means continuity planning cannot stop at API uptime; it must include regulatory and safeguard changes.

The case also exposes a difficult threshold problem. A model that assists authorized vulnerability research may provide similar techniques to an attacker. Blocking every dual-use request can weaken defenders, while relying on self-declared intent is inadequate. Anthropic’s proposed industry framework could help if it defines severity, reproducible testing, response times, and disclosure norms in a way outsiders can inspect.

The accurate current status is that Fable 5 is available again, not that it remains restricted during review. The unresolved questions are how the new classifier performs on benign tasks, whether bypass variants emerge, and whether the promised shared framework becomes a public, testable standard.