Anthropic disables Fable 5 and Mythos 5 globally

Anthropic received an export control directive from the United States government to suspend access to its most advanced models, Fable 5 and Mythos 5, for any foreign national, according to Anthropic's public statement dated Jun 12. The company quotes the timing of the notice as 5:21pm (ET) and says, "we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance." Reuters and The Guardian report the directive did not include written specifics of the government's national security concern.

Anthropic's statement and reporting by ITNews describe differences between the two releases: Fable 5 was made available for public access on June 10 and included safeguard layers intended to block use for cybersecurity, biological and chemical research; Mythos 5 was distributed only to select partners under Project Glasswing and reportedly omitted some of the safeguard layers present in Fable 5. Anthropic's statement also says it reviewed a demonstration the government provided that allegedly shows a method to bypass safeguards that can identify a small number of previously known, minor software vulnerabilities.

Anthropic's public post characterises the government's disclosure as "verbal evidence of a potential narrow, non-universal jailbreak." The company is quoted as writing, "We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people," language repeated in Reuters and The Guardian coverage. ITNews reports Anthropic worked with US agencies, the UK AI Security Institute (AISI) and third parties to red-team Fable 5 prior to launch, and that the company applied a defence-in-depth approach and a 30-day data retention window for usage monitoring.

Public reporting frames this action as a notable expansion of export-control tactics from hardware and development tools to the models themselves. Observers will see this as an example of regulators using national-security authorities to limit foreign access to deployed models, rather than only restricting compute chips or software components. For practitioners, the episode highlights that access to commercial LLMs can be interrupted rapidly by government directives, which complicates risk assessments for deployments that assume continuous model availability.

Companies releasing frontier models have increasingly relied on multi-party red-teaming, layered safeguards and monitoring to manage misuse risks. Industry-pattern observations note that those measures can reduce many classes of misuse but do not guarantee universal resistance to all bypass techniques; public reporting and Anthropic's own language acknowledge imperfect jailbreak resistance. The incident is likely to accelerate policy discussions about how governments evaluate and, if necessary, intervene on deployed AI capabilities.

Follow whether the US government provides a written disclosure or technical report describing the alleged jailbreak technique, and whether other vendors receive similar directives. Observers should also track customer impact notices from cloud and platform partners, and any regulatory or legal challenges that arise. Finally, watch for public releases from security researchers or agencies that corroborate or refute the specific bypass demonstration described in coverage.