The US Commerce Department lifted export curbs on Anthropic's two most advanced models on July 1, restoring global access to Claude Fable 5 and returning Claude Mythos 5 to US organizations three weeks after the administration flagged both as national security risks. Access to Mythos 5 was reinstated for domestic users on June 26, and Anthropic said it is now working with the government to expand Mythos into a broader set of partners under its Glasswing cybersecurity program.
Commerce Secretary Howard Lutnick told Anthropic in a letter that the company would no longer need a license for exports or in-country transfers of the Mythos and Fable models. The letter credited Anthropic with taking steps in close coordination with the government to address the risks, while reserving the right to reimpose curbs at any point. Lutnick said the two weeks of joint analysis were meant to ensure alignment across the US Government.
The June 12 shutdown order came from fears that China, Russia, or other countries of concern could exploit the models to attack US infrastructure, including the electric grid and banking system. Anthropic responded by cutting off all access because it lacked a way to block users by country. Mythos in particular was described by the company as uniquely attractive to malicious actors, capable of finding and exploiting software vulnerabilities more effectively than any other model and all but the most skilled human security experts.
“Over the past two weeks, we have worked closely with Anthropic to analyze and approve Fable 5 to ensure alignment across the US Government and strengthen America's leadership in AI”— Howard Lutnick, US Commerce Secretary
Key facts
- 01The US Commerce Department lifted export curbs on Claude Fable 5 and Claude Mythos 5 on July 1, 2026, roughly three weeks after imposing them.
- 02US organizations regained access to Mythos 5 on June 26; Fable 5 is now available globally.
- 03Anthropic said the jailbreak method flagged by Amazon researchers is now blocked in over 99% of cases on Fable 5.
- 04A dedicated internal team will monitor jailbreak reports 24/7, alongside a new HackerOne program for security researchers.
- 05Anthropic is working with Amazon, Microsoft, and Google to draft a four-criteria framework for scoring jailbreak severity.
Fable 5 shares the same underlying model as Mythos 5 but, according to Anthropic, provides none of those unique offensive capabilities. The trigger for the June curbs was a bypassing method discovered by Amazon researchers that manipulated Fable 5 into producing code demonstrating how a vulnerability could be exploited. Anthropic's own testing found that less advanced rival models, including GPT-5.5 and Kimi K2.7, could identify the same vulnerabilities, which the company said confirmed the technique did not expose any unique Mythos-level cyber capabilities and only involved routine defensive cybersecurity work.
The company moved to close the gap anyway. Anthropic said the reported jailbreak is now blocked in over 99 percent of cases, achieved by training a new safety classifier in coordination with the government. The trade-off is that some benign prompts may be blocked during routine coding and debugging tasks.
When Fable 5 refuses a request, the user is notified and the query is rerouted to Opus 4.8. Anthropic acknowledged that the new classifier can make mistakes and has long maintained that building a model fully impervious to jailbreaks is probably impossible. The goal, the company said, is to make successful jailbreaks costly and high-effort, and to catch them before malicious actors can weaponize them.
“Users will be notified if a request to Fable 5 is blocked, and the request will instead be sent to Opus 4.8”— Anthropic, company blog post
To operationalize that, Anthropic set up a dedicated internal team that monitors jailbreak submission channels 24/7 and launched a new HackerOne program for outside security researchers to submit potential cyber jailbreaks in Fable 5. It is also working with Amazon, Microsoft, and Google, alongside other Glasswing partners, to draft a consensus framework for assessing jailbreak severity. The framework establishes four scoring criteria: how much capability a jailbreak provides, how many offensive tasks it enables, how easy it is for a human to weaponize, and whether it requires specialist knowledge to discover.
The episode has deepened Anthropic's ties to the US government, a notable shift given that the company recently sued the government over a national security risk designation. Anthropic had argued the designation was retaliation after it refused to grant access to its models for building autonomous weapons or conducting domestic mass surveillance. It is now expanding its pre-deployment testing commitments, offering the government early access to frontier models and dedicating resources to joint AI-evaluation research.
White House Chief of Staff Susie Wiles framed the outcome as a validation of the administration's approach, saying the government and private sector have worked together in a way we have never seen before and that the shared priority remains getting the best tech deployed as quickly and safely as possible. Anthropic's blog urged Congress to pass laws that would hold all frontier model developers to the same standard, arguing the current arrangement is a template for global coordination but only a beginning.
The unresolved piece is speed. Two people familiar with the discussions said the administration is still working on a framework for how companies should formally submit new AI models for review and what standards they would be held to. CEO Dario Amodei made the point earlier this month with a Lord of the Rings reference, comparing the pace of political institutions to Treebeard the Ent — wise but slow enough that the forest could be cut down before he acts.
Not every risk raised by Amazon's finding is fully closed off. Anthropic said the attack still works in a very small fraction of cases, though in those cases the model may provide information that isn't detailed enough to help a cyberattacker. The company's layered defenses, which include the trade-off of blocking some benign requests, are meant to catch what the classifier misses. Anthropic did not commit to a timeline for eliminating the residual failure mode.
For the AI market, the three-week round trip is the template now. A frontier lab ships a capable model, a government flags a national security concern, the model is pulled, safeguards are retrained, and access returns with tighter classifiers and a formalized red-team pipeline. Anthropic emerged with expanded government contracts and a co-authored jailbreak-scoring framework its rivals will effectively have to adopt. Competitors watching this play out — particularly the labs sitting on models with comparable offensive-security capabilities — should expect the same review to arrive at their door, and should assume the price of clearance is deeper federal integration rather than lighter oversight.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
