Anthropic has restored access to its Claude Fable 5 and Claude Mythos 5 AI models after the US government lifted export controls imposed earlier this month. Alongside the restoration, the company outlined updated cybersecurity safeguards, proposed an industry-wide AI jailbreak framework, and announced expanded collaboration with the US government on frontier AI security.
The temporary suspension began on June 12 after the US government restricted access to Anthropic’s newest AI models over cybersecurity concerns.
Why Fable 5 and Mythos 5 were suspended
Anthropic said the June 12 export controls required it to restrict access for foreign nationals, regardless of whether they were inside or outside the United States. Because the order took effect immediately and Anthropic lacked a reliable way to verify nationality in real time, it suspended access to both models for all users.
Claude Fable 5 and Claude Mythos 5 were launched on June 9 and share the same underlying model architecture. However, Fable 5 was released with extensive safeguards for broader use, while Mythos 5 was limited to trusted Project Glasswing partners for defensive cybersecurity applications.
Amazon report and updated safeguards
Anthropic said the export controls followed a report from Amazon researchers who discovered a method of bypassing some of Fable 5’s cybersecurity safeguards. The reported technique allowed the model to identify software vulnerabilities and, in one instance, generate code demonstrating how a vulnerability could potentially be exploited.
Following an investigation conducted with the US government, Amazon, and other partners, Anthropic found that:
- Claude Opus 4.8, GPT-5.5, and Kimi K2.7 could identify the same vulnerabilities.
- Claude Haiku 4.5, Sonnet 4.6, Opus 4.6, Opus 4.7, Opus 4.8, GPT-5.4, GPT-5.5, and Kimi K2.7 could produce similar exploit demonstrations.
- The reported bypass did not expose any unique Mythos-level cybersecurity capabilities.
- The incident involved a borderline case of routine defensive cybersecurity work that had been blocked out of caution.
Anthropic subsequently developed an updated safety classifier to block the reported bypass technique. According to the company, the updated safeguards:
- Block the reported technique in more than 99% of cases.
- Redirect blocked requests to Claude Opus 4.8.
- May increase false positives during routine coding and debugging tasks.
- Were tested by researchers from the US Department of Commerce’s Center for AI Standards and Innovation (CAISI), which agreed that the safeguards are “extraordinarily strong.”
How Anthropic’s cybersecurity safeguards work
Anthropic said Claude Mythos 5 can identify and exploit software vulnerabilities more effectively than any other AI model and all but the most skilled human security experts, making it particularly attractive to malicious actors, according to the company.
Anthropic also said Claude Fable 5 was launched with its strongest cybersecurity safeguards to date and that it doubled the number of researchers and engineers working on these protections before launch.
Fable 5 uses a “defense in depth” approach that combines:
- Model training to refuse dangerous requests
- Automated safety classifiers
- Misuse monitoring systems
- Retrospective analysis of model behavior
- Additional layered safeguards
Anthropic said safety classifiers continuously evaluate prompts and outputs to identify potentially harmful cybersecurity activity. The company acknowledged that classifiers can occasionally miss harmful requests, block legitimate requests, or be bypassed through jailbreak techniques.
To reduce these risks, Anthropic expanded Fable 5’s “safety margin,” requiring requests to appear clearly safe before they are approved. Anthropic categorized jailbreaks into three groups: minor jailbreaks that intrude into the safety margin, narrow harmful jailbreaks that enable specific harmful behaviors, and universal jailbreaks that unlock broad harmful capabilities. The company said all reported Fable 5 jailbreaks so far fall into the minor category and that no universal jailbreaks have been discovered.
Proposed AI jailbreak framework
Anthropic said it is working with Amazon, Microsoft, Google, and other Project Glasswing partners to develop a common framework for assessing AI jailbreak severity.
The proposed framework evaluates jailbreaks using four criteria:
- Capability gain
- Breadth of capability gain
- Ease of weaponization
- Discoverability
According to Anthropic, the framework could help companies and governments determine the severity of jailbreaks and coordinate responses. For the most severe cases, such as attacks affecting critical infrastructure or financial systems, the company said it would deploy mitigations immediately after confirming the threat.
Anthropic also plans to establish a dedicated 24/7 monitoring team for jailbreak reports and launch a HackerOne program for cybersecurity jailbreak submissions involving Claude Fable 5.
Expanded US government collaboration
Anthropic said it has worked closely with the US government over the past ten weeks as policymakers developed the June 2 Executive Order on Promoting Advanced Artificial Intelligence Innovation and Security. The company’s engagement included the Office of the National Cyber Director, the Office of Science and Technology Policy, the Department of the Treasury, the Department of Commerce, CAISI, and other national security agencies.
Building on nearly two years of existing collaboration, Anthropic announced commitments that include:
- Expanded pre-release access for government testing of frontier AI models and safeguards
- Faster information sharing on significant jailbreaks and misuse patterns
- Dedicated teams, compute resources, and safety expertise for joint AI security research
- Participation in the interagency cybersecurity vulnerability clearinghouse established under the June 2 Executive Order
- Cooperation with governments and industry partners on shared security and evaluation standards
Anthropic said it hopes these efforts will contribute to broader international coordination on advanced AI safety and security and argued that future AI safety requirements should eventually be codified through stronger regulation and applied consistently across frontier AI developers.
Availability
Anthropic said Claude Fable 5 is now available globally through Claude Platform, Claude.ai, Claude Code, and Claude Cowork.
- Pro, Max, Team, and select Enterprise users will receive access to Fable 5 for up to 50% of their weekly usage limits through July 7.
- After July 7, access will move to a usage credit model.
- Anthropic plans to restore access through Amazon Web Services, Google Cloud, and Microsoft Foundry as quickly as possible.
- Claude Mythos 5 access has been restored for a set of US organizations following government approval on June 26.
- Anthropic continues to work with the US government to expand access to additional domestic and international Project Glasswing partners.