Anthropic Restores Claude Fable 5 Globally After US Lifts Export Controls, While Mythos Access Remains Limited

Anthropic has restored global access to Claude Fable 5 after the United States government lifted export controls that forced the company to suspend its newest artificial intelligence models earlier in June. Claude Mythos 5 has also returned, although its availability remains limited to a selected group of approved United States organizations while Anthropic negotiates wider access for international cybersecurity partners.

According to Anthropic’s official redeployment announcement, the restrictions on Fable 5 and Mythos 5 were lifted on June 30, 2026. Fable 5 became available globally again on July 1 through Claude.ai, the Claude Platform, Claude Code, and Claude Cowork. Anthropic is also working to restore access through AWS, Google Cloud, and Microsoft Foundry. The latter recently became an important part of Anthropic’s infrastructure strategy, as Claude running through Microsoft Azure on NVIDIA GB300 Blackwell Ultra.

The restoration follows an unusual sequence of government intervention. On June 12, the United States government issued an export control directive requiring Anthropic to prevent every foreign national from accessing Fable 5 and Mythos 5, regardless of whether that person was located inside or outside the United States. The restriction also applied to foreign national employees working directly for Anthropic.

Anthropic received the directive at 5:21 PM Eastern Time and said the accompanying letter cited national security authorities without identifying the specific technical concern. Because the order took effect immediately and Anthropic lacked a reliable system for confirming every user’s nationality in real time, the company suspended both models for everyone rather than risk violating the directive.

The company initially criticized the action as overly broad and technically unsupported. Anthropic said the government had only provided verbal information about a narrow method of bypassing Fable 5’s safeguards and had not disclosed evidence of a universal jailbreak capable of broadly unlocking the model’s restricted cybersecurity capabilities.

Anthropic has now provided more information about what triggered the intervention. The government reportedly became aware of research from Amazon showing that Fable 5 could be prompted to identify several software vulnerabilities despite its safeguards. In one case, the model also generated code demonstrating how a vulnerability could be exploited.

Anthropic said its investigation found that the same vulnerabilities could be identified by several less capable models. The company also claimed that the relevant exploit demonstration could be reproduced using earlier Claude models, GPT 5.4, GPT 5.5, and Kimi K2.7. Anthropic therefore concluded that the report did not expose unique Mythos level offensive capabilities, although it still developed additional protections in response.

The new safety classifier is designed to recognize and block the specific technique described in Amazon’s report. Anthropic says the updated system blocks the behavior in more than 99% of tested cases. When a Fable 5 request is flagged, the user will be informed and the request will instead be processed by Claude Opus 4.8.

Anthropic acknowledged that the stronger classifier will temporarily create more false positives during ordinary coding, debugging, and defensive security work. The company plans to continue refining the system so that it can distinguish harmful activity from legitimate software development with greater precision.

Fable 5 and Mythos 5 share the same underlying model, but they are deployed under substantially different conditions. Fable 5 includes Anthropic’s strongest safety systems and is intended for broad commercial and consumer use. Mythos 5 includes fewer restrictions and provides significantly stronger access to advanced cybersecurity capabilities, which is why it remains limited to trusted organizations participating in Project Glasswing.

Anthropic has restored Mythos 5 access to a group of United States organizations following government approval granted on June 26. However, the wider group of domestic and international Glasswing partners has not yet regained complete access. Anthropic says it is continuing discussions with the government to expand availability.

Project Glasswing began with approximately 50 organizations and was later expanded toward another 150 partners across more than 15 countries. Participants include companies and institutions responsible for operating systems, browsers, cloud infrastructure, financial services, communication networks, healthcare platforms, and critical open source software. Many of these organizations use Mythos to scan large codebases, identify vulnerabilities, and develop patches before attackers can exploit them.

The model attracted substantial attention in April when Anthropic revealed that Claude Mythos Preview had identified thousands of high severity vulnerabilities across every major operating system and web browser. The discoveries included a 27 year old vulnerability in OpenBSD, a 16 year old issue in FFmpeg, and multiple Linux kernel vulnerabilities that could be combined to obtain complete control of a system.

Anthropic said many of the vulnerabilities had survived decades of human review and millions of automated security tests. Mythos Preview was also capable of autonomously constructing exploits, including complex chains combining several weaknesses. The company has argued that these capabilities could provide defenders with a major advantage, but could also create severe risks if accessed by criminal organizations or hostile governments.

Anthropic CEO Dario Amodei highlighted those concerns in his June essay on artificial intelligence policy, arguing that the cybersecurity capabilities demonstrated by Mythos prove that frontier models have become strategically important technologies.

“The evidence of AI’s incredible power, as well as its risks, has become undeniable.”
— Dario Amodei

Amodei warned that future models could disrupt financial systems, critical infrastructure, and national security if safeguards and government policy fail to advance quickly enough. Anthropic has supported stronger regulation for highly capable models, including government authority to block deployments that present credible catastrophic risks. However, the company has also argued that this authority must follow a transparent and technically grounded process.

The Commerce Department dispute is separate from Anthropic’s earlier confrontation with the Pentagon. In March, the company said it had been designated a supply chain risk following disagreements about the use of Claude for fully autonomous weapons and mass domestic surveillance. Anthropic challenged that designation and maintained that its limitations were necessary because existing AI systems were not reliable enough for those uses.

The latest resolution appears more cooperative. Researchers from the Commerce Department’s Center for AI Standards and Innovation tested Anthropic’s original and updated safeguards, while Anthropic agreed to deeper collaboration with the government on model testing, information sharing, and cybersecurity research.

Anthropic is also working with Amazon, Microsoft, Google, and other Project Glasswing members to create a shared framework for measuring jailbreak severity. The proposed system would help companies and regulators distinguish between harmless bypasses, narrow harmful failures, and universal jailbreaks that provide broad access to restricted capabilities.

The temporary removal of Fable 5 and Mythos 5 demonstrates how difficult frontier AI regulation will become as models develop capabilities with both defensive and offensive applications. A system capable of discovering vulnerabilities can help protect hospitals, banks, operating systems, and power infrastructure, but the same capability can assist attackers if safeguards fail.

The government’s initial directive created a major transparency problem. Anthropic was ordered to remove access immediately without receiving a detailed technical explanation, forcing the company to disable the models for every customer. That approach protected the government from a possible security risk, but it also removed an important defensive tool from trusted organizations around the world.

Anthropic’s response also deserves scrutiny. The company frequently emphasizes the extreme capabilities and possible dangers of its models while simultaneously arguing that specific safeguards make broad deployment acceptable. That position creates tension between commercial expansion and the company’s warnings about national security risks.

The improved classifier and proposed jailbreak framework are constructive steps, but neither eliminates the underlying challenge. Anthropic itself acknowledges that perfect resistance to jailbreaks may not currently be possible. Regulators and AI companies will therefore need systems that respond proportionally to specific evidence rather than using complete model shutdowns as the default solution.

Should governments have the authority to immediately suspend powerful AI models, or should companies receive a transparent technical review before access is restricted?