Anthropic restores Claude Fable 5 after US export controls are lifted

Anthropic says access to Claude Fable 5 and Claude Mythos 5 has been restored after a short but significant interruption caused by US export controls. The company explains that on June 12, 2026, the US government applied immediate restrictions to the two new models. Because Anthropic could not reliably verify each user’s nationality in real time, it temporarily suspended access for everyone rather than risk violating the order. The controls were later lifted for Fable 5 and Mythos 5, allowing Anthropic to begin bringing the models back online.

Fable 5 is being returned to global users across Anthropic’s main products, including the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. For Pro, Max, Team, and selected Enterprise customers, the model will count for up to half of weekly usage limits through July 7, after which it will be available through usage credits. Anthropic also says it is working to restore access through AWS, Google Cloud, and Microsoft Foundry as soon as possible, which matters because many companies use Claude through cloud providers rather than directly through Anthropic’s own apps.

Mythos 5 is being handled more cautiously. Anthropic says access has been restored only for a set of US organizations after government approval on June 26. The company is still coordinating with the government to expand availability to more domestic and international partners in Project Glasswing. In the source text, Mythos 5 is described as sharing the same underlying model as Fable 5, but with fewer safeguards and a narrower release path. It was intended for trusted partners working on defensive cybersecurity, not for broad public use.

The incident began shortly after the June 9 launch of both models. According to Anthropic, the export control directive followed a report from Amazon researchers who found a way to bypass Fable 5’s safeguards. The reported prompt technique made the model identify software vulnerabilities, and in one case it generated code showing how a vulnerability could be exploited. In simple terms, this was a “jailbreak”: a way of phrasing a request so that an AI system does something its safety rules are meant to prevent or limit.

Anthropic says it reviewed the report with the US government and partners including Amazon. Its own testing found that the same vulnerability-identification behavior was not unique to Fable 5. The company says several less capable models could identify the same vulnerabilities, and that every model it tested could produce the same exploit demonstration for the single vulnerability highlighted in the report. Anthropic lists models including Claude Haiku, Sonnet and Opus variants, GPT-5.4, GPT-5.5, and Kimi K2.7 as being able to show similar behavior in that specific case.

Anthropic argues that the report did not reveal unique Mythos-level cyber capabilities, but rather a borderline failure in Fable 5’s safeguards. The company frames the behavior as routine defensive cybersecurity work that had been blocked out of extra caution. That distinction is important: Anthropic is not saying the bypass was irrelevant, but it is saying the issue did not prove that Fable 5 had crossed into a new category of dangerous cyber capability. Still, the company says it moved quickly to address the reported bypass because even borderline safety failures can become serious when highly capable models are deployed at scale.

To fix the issue, Anthropic trained a new safety classifier aimed specifically at the behavior described in the Amazon report. A safety classifier is a system that checks requests or outputs and decides whether they should be allowed, blocked, or redirected. Anthropic says the new classifier blocks the reported jailbreak technique in more than 99% of cases. When a user’s request to Fable 5 is blocked, the user will be notified and the request will instead be routed to Opus 4.8. The company adds that in a very small fraction of cases, the model may still provide information, but not at a level of detail that would meaningfully help a cyberattacker.

The post also uses the incident to call for a broader industry framework for evaluating jailbreaks. Anthropic says AI companies need a shared way to assess how severe a jailbreak is, how quickly it must be fixed, and how to communicate its risk to governments and partners. Without a common standard, one company’s “low-risk bypass” could be another company’s “critical vulnerability,” making coordinated response harder. Anthropic says it has started developing such a framework with Amazon, Microsoft, Google, and other Glasswing partners.

The larger message is that frontier AI deployment is becoming tied to government review, cloud distribution, and cybersecurity risk management. Anthropic presents the resolution as constructive, but the episode shows how quickly access to major AI models can be interrupted when regulators see a potential security issue. It also shows that AI safety is no longer only about model behavior inside a product. It now includes pre-release testing, information sharing with governments, coordinated partner review, and clearer rules for deciding when a model’s cyber capabilities are acceptable, risky, or too dangerous to release broadly.