Anthropic’s most advanced AI models are back. The US lifted export restrictions on Fable 5 and Mythos 5 about three weeks after the Trump administration flagged them as national security risks.
Fable 5 is available globally starting today. US organizations got Mythos 5 back on June 26. Anthropic is also working to expand Mythos access to “a broader set of domestic and international partners” through the Glasswing program — that’s the initiative letting cybersecurity researchers at trusted companies use Mythos for defense work.
Commerce Secretary Howard Lutnick confirmed in a letter that Anthropic no longer needs a license to export or transfer the models. He acknowledged the company worked with the government to address the risks.
So what changed? Anthropic agreed to expand its government partnership. It set up a program with hackers to red-team the models and created a dedicated internal team to monitor jailbreak threats around the clock.
Lutnick was clear: the US can reimpose curbs at any point if it changes its mind. But for now, both he and White House Chief of Staff Susie Wiles are celebrating the deployment.
The whole thing started when Amazon researchers found a bypass method in Fable 5 that could produce code showing how to exploit vulnerabilities. Anthropic says testing confirmed that rival models like GPT-5.5 and Kimi K2.7 could find the same bugs. In their view, the Amazon report didn’t expose anything unique to Mythos-level capabilities — it was “routine defensive cybersecurity work.”
Still, Anthropic tightened the safeguards. The company trained an improved safety classifier that blocks the reported bypass in over 99% of cases. The trade-off: some benign prompts may get blocked during routine coding. When that happens, the request gets routed to Opus 4.8 instead.
Anthropic admits it’s “probably impossible” to build a model fully immune to jailbreaks. The goal is to find them first. The company is now working with Amazon, Microsoft, Google, and other Glasswing partners to create a consensus framework for scoring jailbreak severity — four criteria for determining how bad a break is and what the response should be.
Other industry partners are welcome to join those talks. The whole episode shows how quickly the line between national security and AI capability is blurring.
