Anthropic restores Claude Fable 5 & Mythos 5 access

Anthropic has restored access to its Claude Fable 5 and Mythos 5 models after the US Department of Commerce lifted export controls. The restrictions had forced the company to suspend both models for all users.

Fable 5 will return to global users across Claude products. Mythos 5 had already been restored for a group of US organisations under separate government approval.

The decision reverses an abrupt shutdown that began when US authorities imposed export controls on the two models over national security concerns tied to cybersecurity safeguards. Anthropic initially restricted access for foreign nationals under the order. But because it could not verify nationality in real time, it suspended both models entirely.

The government action followed a report from Amazon researchers who found a way to bypass Fable 5's safeguards. Anthropic said the method prompted the model to identify several software vulnerabilities and, in one case, produce code showing how a vulnerability could be exploited.

Anthropic's internal review found that less capable models could identify the same vulnerabilities, and that every model it tested could produce the same exploit demonstration. It added that the technique did not reveal any unique Mythos-level cyber functions.

The dispute centred on the distinction between Anthropic's two newest models. Fable 5 was released for general use with tighter restrictions intended to block dangerous cybersecurity requests. Mythos 5 was made available only to a limited group of trusted partners for defensive cybersecurity work.

Safeguard changes

To secure the lifting of the controls, Anthropic worked with the government and partners, including Amazon, to revise its protections. It trained an improved safety classifier to block the behaviour described in the report and will redirect blocked Fable 5 requests to an older model, Opus 4.8.

Anthropic said the updated classifier blocked the specific technique described in the Amazon report in more than 99% of cases. It also warned that the added safeguard increased the risk of false positives during routine coding and debugging work.

Researchers at the Department of Commerce's Centre for AI Standards and Innovation tested both the previous and revised protections, according to Anthropic. Those researchers agreed the safeguards were "extraordinarily strong", it said.

The episode highlights tensions between rapid model releases and government efforts to prevent misuse of advanced AI systems in cyberattacks. Anthropic argued that Fable 5 did not offer unique offensive cyber advantages because of the restrictions built into the model. Mythos 5 remained under tighter controls because of its broader utility in identifying and exploiting software vulnerabilities.

Anthropic also used the reinstatement to outline a broader push for common standards on so-called jailbreaks, prompting methods used to bypass model restrictions. It said there was no industry-wide agreement on how to measure the seriousness of a jailbreak, creating uncertainty for companies and regulators when new techniques emerge.

Industry framework

Anthropic has begun work with Amazon, Microsoft, Google and other Glasswing partners on a framework to assess jailbreak severity. The proposal would score findings on capability gain, breadth of capability gain, ease of weaponisation and discoverability.

It said the framework could help determine how quickly a company should respond when a bypass is discovered. Anthropic also said it was creating a team to monitor key jailbreak submission channels around the clock.

The company also outlined a broader agenda for cooperation with US agencies on frontier AI security. That includes giving designated government partners expanded early access to models and safeguards for pre-release testing, quicker notification when significant jailbreaks or misuse patterns are identified, and dedicated staff and computing resources for joint research.

The restoration of service offers relief for customers after a disruption that affected one of Anthropic's newest commercial launches. Fable 5 had been included for some subscribers within weekly usage limits before moving to a usage-credit model, while access on third-party cloud platforms was still being re-enabled.

Anthropic said the original suspension was broader than the legal order's practical target because it lacked a reliable way to distinguish foreign nationals from other users at the point of access. That limitation turned a control aimed at a specific category of users into a full shutdown of both models.

Anthropic maintained that the government's earlier action was based on a narrow finding rather than evidence of a broadly effective jailbreak. It also said no universal jailbreak for Fable 5 had been found, while acknowledging that perfect resistance to jailbreaks was unlikely for any model provider.

"Access to Claude Fable 5 and Mythos 5 is now restored."