Key Summary
- On the evening of June 12th, the U.S. Department of Commerce issued a letter prohibiting foreigners from accessing the devices, forcing Anthropic to shut down Fable 5 and Mythos 5 to its global customers.
- The trigger was a third party's claim of "jailbreaking" Mythos 5, but Anthropic refuted this, stating that the vulnerability was already known and could be exploited even with GPT-5.5.
- For the first time, the Ministry of Commerce has applied export control licensing authority to the AI model itself, requiring licenses to be applied for on a case-by-case basis for both export and transfer.
The US government's move was swift and decisive. At 5:21 PM on June 12, Anthropic received an official letter from the US Department of Commerce, demanding an immediate ban on any foreigners accessing Fable 5 and Mythos 5, which had only been online for three days.
The letter was signed by Commerce Secretary Howard Lutnick and drafted by officials from the Commerce Department's Bureau of Industry and Security (BIS). To comply with regulations, Anthropic had no choice but to disconnect both models from "all" customers, including its own foreign employees.
From its launch to the global shutdown, it only took three days.
However, this sudden halt had actually been brewing for several weeks. According to Axios, citing a government official, even before the two models were officially released on June 9, the Department of Commerce was deeply concerned about the models' "extremely strong cybersecurity inference capabilities" and privately tried several times to persuade Anthropic to postpone the release, but without success.
Anthropic adopted a compromise at the time. The publicly released version was the commercial Fable 5, equipped with extremely stringent filters. When encountering questions related to cybersecurity, biology, or chemistry, it would automatically transfer the conversation to the less capable Claude Opus 4.8. The truly "unlocked" bare-bones model, Mythos 5, was kept under the tight control of "Project Glasswing," accessible only to about 50 background-checked red team organizations, critical infrastructure operators, and defense departments.
This concern is not unfounded. Anthropic itself revealed that in the first few weeks after the Glasswing project went live, Mythos, with 50 partners, discovered more than 10,000 high-risk and critical vulnerabilities in important software. During testing, the model could find and exploit zero-day vulnerabilities in every major operating system and every major browser, as long as the user gave the command.
The rumored reason was a "prison break".
What truly ignited the fuse was that just three days after the model's release, a third-party company claimed to have successfully "jailbroken" the controlled Mythos 5 and reported the results. For Lutnik and the Ministry of Commerce, this was the perfect opportunity for a strong intervention.
However, Anthropic's version is different. The company says that the demonstration it reviewed was actually a filter for the public version of Fable 5. The method is to have the model read a specific piece of code, find and fix software defects, and finally only find "a few minor vulnerabilities that were already known." The method is quite simple, and even other public models such as OpenAI's GPT-5.5 can do it.
The export control first-time restriction model itself
This time, the Ministry of Commerce has invoked the "export control" licensing authority previously used to regulate chips and weapons. According to the official letter, Anthropic will need to apply for a license on a case-by-case basis for any future export, re-export, or transfer of these models within the country; violations will face civil and monetary penalties. This unprecedented step of using the export control stick, originally wielded against physical goods, to directly target a software model is truly groundbreaking.
- The ban applies to all foreign nationals both inside and outside the United States, including Anthropic's foreign employees; all are prohibited from accessing [the service/resources].
- The licensing requirements apply to export, re-export, and inland transfer models, each requiring a license application on a case-by-case basis.
- Violations will result in civil and monetary penalties.
- Only Fable 5 and Mythos 5 are affected; other Anthropic models remain unaffected.
Anthropic disagrees, issuing a statement emphasizing that "we believe this is a misunderstanding and are working to restore access as quickly as possible." The company also stated bluntly that if a commercially viable model deployed to hundreds of millions of people is recalled due to a potentially narrow jailbreak, applying this standard to the entire industry would essentially bring all new model deployments by leading model vendors to a complete standstill.
We disagree that discovering a narrow potential jailbreak should be a reason to recall a commercially available model that has been deployed to hundreds of millions of people.
An official revealed that these two models may be locked up for "the next few weeks" and will not be lifted until the government strengthens its national security measures. Ironically, just one day before the ban was imposed, Anthropic CEO Dario Amodei publicly argued that "the government should have the right to veto high-risk AI." Now the government has indeed intervened, but not in the way he wanted.
Frequently Asked Questions
Why did the US government freeze Anthropic's Fable 5 and Mythos 5?
On June 12, Commerce Minister Lutnik issued a letter citing national security reasons, invoking export control licensing authority to prohibit foreigners from accessing the device. The trigger was a third-party company's claim of successfully jailbreaking the Mythos 5. To comply with regulations, Anthropic had no choice but to directly shut down these two models to all its global customers.
What are the differences between Fable 5 and Mythos 5?
Fable 5 is a publicly available commercial version with strict filters. Questions about cybersecurity and biohazards are automatically forwarded to Claude Opus 4.8 for processing. Mythos 5, on the other hand, is a bare model with the security lock removed, and is only made available to about 50 approved red team organizations and defense units through the "Glass Wing Project".
Related reports
Uniswap founder criticizes Claude Fable 5's "security filter" as stupid, skipping essential checks.
