Fable 5 Restrictions Lifted After 18 Days, Anthropic Details Safety Measures

Anthropic announced on June 30 that it has resumed global access to its large language model, Claude Fable 5, following the U.S. Department of Commerce’s decision to lift export restrictions imposed on June 12. The decision was revealed on the company’s official blog.

The export restrictions, which lasted 18 days, were resolved by the implementation of a targeted safety filter designed to block a specific prompt technique identified by Amazon researchers. The U.S. Department of Commerce’s Center for AI Standards and Innovation (CAISI) reviewed the safety measures before withdrawing the restrictions.

Origins and Timeline of the Issue

On June 12, the U.S. Department of Commerce enacted export restrictions on Fable 5 and its underlying model, “Mythos 5.” The restrictions applied to usage by non-U.S. nationals, including non-U.S. employees of Anthropic. Due to its inability to determine user nationality, Anthropic decided to suspend access to both models worldwide.

The crisis began when Amazon researchers reported that they had discovered a specific prompt technique that could manipulate Fable 5 into identifying software vulnerabilities and generating exploitative code for those vulnerabilities.

Implemented Safety Measures

To address the issue, Anthropic trained a new classifier capable of blocking the specific prompt technique reported by Amazon researchers with an accuracy of over 99%. Requests flagged by this classifier are redirected to an older version of the model, Opus 4.8. Anthropic has acknowledged that this solution may inadvertently block legitimate coding and debugging requests.

Importantly, the classifier targets the reported prompt technique without diminishing the model’s overall capabilities. Fable 5 remains capable of identifying the vulnerabilities described in Amazon’s report. The classifier’s approach involves redirecting flagged requests rather than reducing the model’s functionality.

However, detection-based safety measures, like the one implemented, are inherently limited. While the classifier effectively addresses known techniques, it remains vulnerable to yet-undiscovered methods. Anthropic has acknowledged that no model can be fully jailbreak-proof and anticipates that additional jailbreak methods will emerge in the future.

Current Status of Fable 5

Fable 5 is now accessible via Claude.ai, Claude Platform, Claude Code, and Claude Cowork. The model is also expected to become available again on AWS, Google Cloud, and Microsoft Foundry in the near future.

Anthropic’s investigation, conducted in collaboration with the U.S. government and Amazon, revealed that other models, including Opus 4.8, OpenAI’s GPT-5.5, and China’s Kimi K2.7, were also capable of identifying the same vulnerabilities and generating exploitative code. All tested models, including Haiku 4.5, Sonnet 4.6, and multiple versions of Opus, were able to replicate the exploit. These findings call into question the previously overstated cybersecurity capabilities of the Mythos class.

The return of Fable 5 has allowed it to reclaim the top spot on the benchmark leaderboards, a position held by China’s Z.ai GLM-5.2 during Fable 5’s absence. Fable 5 has again achieved the top score in the multi-week task test, “AA-Briefcase.”

Meanwhile, Mythos 5, which operates with fewer safety constraints and is accessible only to Project Glasswing partners, was partially reintroduced to select U.S.-based organizations on June 26.

Future Vulnerability Reporting Framework

Anthropic has also launched a HackerOne program to encourage researchers to report new jailbreak methods for Fable 5. Additionally, the company has committed to providing select government partners with access for ongoing audits.

This incident highlights the complex relationship between AI safety and international regulations. The fact that a single prompt technique led to an 18-day global service suspension underscores the challenges of AI governance.

Editorial Opinion

In the short term, the lifting of restrictions allows Fable 5 to regain its competitive edge in the global market. Considering the recent dominance of China’s GLM-5.2 on the benchmark leaderboards during Fable 5’s absence, this is a significant comeback for Anthropic. However, relying on a single classifier to address vulnerabilities introduces the risk of recurrence as new jailbreak methods are discovered. This approach is far from a permanent solution.

In the long term, the process of imposing and lifting export restrictions could redefine the relationship between the U.S. government and AI companies regarding safety standards. The establishment of a review process by CAISI for safety measures may become a standard procedure for regulatory decisions. However, the direct impact of such national regulations on the global availability of AI models raises concerns about the broader goal of democratizing technology.

The editorial team believes that this incident has underscored the limitations of detection-based safety measures, suggesting that the industry should focus on enhancing the intrinsic safety of AI models through architectural improvements.

References

• Tom’s Hardware — Published on 2026-07-01T11:30:51.000Z