The U.S. government removed export restrictions on Anthropic’s latest AI models, Claude Fable 5 and Claude Mythos 5, after previously flagging them as national security risks. Fable 5 is now available worldwide, while U.S. organizations regained access to Mythos 5 on June 26, according to a company announcement.
Commerce Secretary Howard Lutnick informed Anthropic in a letter that the models no longer require export licenses. The decision followed the company’s work to address security concerns, including expanded red-teaming and a 24/7 monitoring team for jailbreak threats.
Lutnick cautioned that restrictions could return at any time. White House Chief of Staff Susie Wiles called the move a success for public-private collaboration under the Trump administration. “Our shared priority remains: deploy the best technology quickly and safely,” she stated.
The original ban stemmed from concerns that adversaries like China or Russia might use the models to target U.S. infrastructure. Mythos 5 was considered especially risky, as Anthropic described its ability to identify and exploit software vulnerabilities with near-expert precision.
Related: Journal retracts two Max Planck papers from 1940s
Fable 5, built on the same architecture, lacks those offensive features. The company added stronger safeguards, though some legitimate coding tasks may still trigger blocks. Flagged requests will redirect users to Opus 4.8.
The policy change came after Amazon researchers found a jailbreak method that could trick the models into revealing exploitable code. Anthropic noted the issue was uncommon, affecting under 1% of cases, and that competing models like GPT-5.5 and Kimi K2.7 could detect the same flaws. A new safety classifier now prevents the attack in over 99% of attempts.
Anthropic admitted no model is fully immune to jailbreaks. It is collaborating with Amazon, Microsoft, and Google to create a framework for evaluating jailbreak severity. The system would assess factors like an attack’s potential damage, ease of use, and required skill.
A HackerOne program now lets security researchers report vulnerabilities in Fable 5. The effort aligns with industry moves to prioritize threat testing and rapid responses.
The government’s reversal signals a change in its approach to Anthropic. The company had sued over a national security designation it claimed was retaliation for refusing to provide models for autonomous weapons or surveillance. Now, it is expanding pre-deployment testing with government partners, offering early access to advanced models and joint research on AI evaluation.
Related: Galaxy S27 Ultra Snapdragon Chip: Another Hint Just Dropped!
CEO Dario Amodei pushed Congress to pass legislation standardizing safety rules. He warned that delays could harm national strategy, comparing the government’s slow pace to a scene from The Lord of the Rings. “The overlap between AI and our political system feels like the Hobbits meeting Treebeard,” he wrote.
Trump initially opposed AI regulations to avoid slowing innovation, but Mythos 5’s release led to a policy shift. His administration is now creating a voluntary framework for safety testing of advanced models, though specifics remain vague.
Isaac Harris, head of the Frontier Security Institute, raised concerns about how the U.S. will manage equally dangerous capabilities from China. Those models often have fewer protections. Anthropic recently accused Alibaba of the largest cloning attack on Claude, urging laws to penalize intellectual property theft by Chinese firms.
Without such steps, Harris warned, adversaries may turn to Chinese alternatives to launch attacks. The situation highlights the challenge of balancing open access with security as AI advances.
