Anthropic has reactivated global access to its Claude Fable 5 AI model, resolving an 18-day disruption triggered by U.S. export restrictions. The decision follows the Department of Commerce’s June 28th withdrawal of controls that had barred non-U.S. nationals from using the model since June 12th, according to a company announcement.
The restoration hinged on a newly deployed safety filter designed to neutralize a specific jailbreak technique identified by Amazon researchers. This classifier, which blocks the identified vulnerability in over 99% of cases, redirects flagged prompts to the older Opus 4.8 model. While effective against the reported exploit, the filter also intercepts some legitimate coding and debugging requests—a trade-off Anthropic acknowledges.
The U.S. Department of Commerce’s Center for AI Standards and Innovation (CAISI) reviewed the safeguards before lifting the export controls. Anthropic confirmed that Fable 5 retains its underlying capability to recognize software vulnerabilities, but the filter now prevents requests that could demonstrate exploits. The company emphasized that detection-based defenses, while targeted at known techniques, remain susceptible to future jailbreak methods.
Why the temporary ban occurred
The export restrictions originated from Amazon’s discovery of a prompt technique that could compel Fable 5 to identify software vulnerabilities and generate exploit code. Anthropic’s classifier specifically targets this method, not the model’s core functionality. However, the incident highlighted broader concerns about cyber-capable AI models, with Anthropic’s own testing revealing that other leading models—including OpenAI’s GPT-5.5 and China’s Kimi K2.7—could also reproduce the exploit demonstration.
Anthropic’s internal review, conducted in collaboration with U.S. authorities and Amazon, found that even smaller models like Haiku 4.5 and Sonnet 4.6 could replicate the single exploit case. These findings reinforced skepticism about claims that certain AI families posed unprecedented cybersecurity risks, particularly in the context of the U.S. government’s sudden restriction on Mythos-class models.
Benchmarks and future safeguards
With Fable 5 back online, Anthropic reclaims its top position on the AA-Briefcase multi-week task test, a benchmark previously held by Z.ai’s GLM-5.2 during the model’s absence. The return also marks a shift in accessibility, as the more capable (and less restricted) Mythos 5 remains limited to select U.S. partners under Project Glasswing.
To proactively address future vulnerabilities, Anthropic has launched a HackerOne program inviting researchers to report new Fable 5 jailbreak attempts. Additionally, the company pledged to grant designated government partners early access to test upcoming frontier models before public release. For existing subscribers, Fable 5 will count toward up to 50% of weekly usage limits through July 7th, after which it will transition to a credit-based system.
The episode underscores the ongoing challenge of balancing AI innovation with security concerns, as regulators and developers navigate an evolving threat landscape where exploit techniques can emerge faster than defensive measures.
AI summary
ABD Ticaret Bakanlığı’nın kaldırdığı ihracat kısıtlamaları sonrası Anthropic’in yapay zeka modeli yeniden erişilebilir hale geldi. Peki, bu süreçte neler değişti ve modelin geleceği ne olacak?
