Why Anthropic has disabled one of its AI models – and what it means
The leading AI company has been ordered to block non-US nationals from accessing its cybersecurity model. The US government’s controversial directive has been met with confusion and unease internationally.
Anthropic wanted to use Mythos to demonstrate what generative AI is capable of in cybersecurity. Three days after the launch of the publicly available version, Fable, both models were taken offline – not because of a technical fault, but on the Pentagon’s orders. The case shows just how much power governments hold over big tech companies and raises questions internationally about dependency.
A «national security risk»
On the evening of 12 June, Anthropic received an export control letter from the Department of Commerce. Under its terms, the company was required to block access to Fable 5 and Mythos 5 for all «foreign nationals», regardless of whether they’re based in the US or abroad. Foreign Anthropic employees were explicitly included in the ban. In practice, the only way the company could meet this requirement was to deactivate the models globally.
The standard Claude models remain online – Fable 5 and Mythos 5 are variants that focus on code analysis and the identification of vulnerabilities. Mythos 5 is the more powerful model, not accessible to the public. Fable 5 features additional safeguards and is more widely available. Anthropic itself had described Mythos as too dangerous to release to the public. Fable 5 reportedly uses a restricted interface to block risky requests relating to exploits and attack tools.
The US government has justified the ban on the basis of evidence that reportedly came in part from Amazon, where researchers had apparently got Fable 5 to identify vulnerabilities in code. While the model initially rejected a direct request, rephrasing it as «fix this code» was enough to get it to return information about bugs that could theoretically be exploited for attacks. The US government argues that this could be used to bypass safety mechanisms, making the model a national security risk.
Anthropic disputes this characterisation. In a public statement, the company describes it as a «potential», «narrow» and «non-universal» jailbreak. According to Anthropic, it only brought to light individual, already known and relatively minor vulnerabilities. It goes on to say that other models such as GPT 5.5 can do the same, sometimes even without bypassing any safety mechanisms. If a case like this is enough to take a model off the market, then according to Anthropic, it would be nearly impossible to productively run any frontier system with code capabilities.
China concerns and a strained relationship
Alongside the alleged jailbreak issue, there were reports that a China-linked group had temporarily gained access to Mythos. From the US government’s perspective, this creates a risk that a mature US model could be reverse engineered through distillation, or partially extracted. There’s currently no evidence for this.
The controversy comes at an already tense time. Anthropic and parts of the US government have been at loggerheads for months: the company has refused to make its models available for mass domestic surveillance and fully autonomous weapons. The Department of Defense subsequently classified Anthropic as a «supply chain risk». The company is taking legal action against this classification. It briefly seemed as though tensions were easing, after agencies sought access to Mythos for cyber-defence projects.
Source: Shutterstock
Experts and politicians baffled by the approach
The case is becoming a test for the US’s new AI strategy, which Trump unveiled in an executive order in early June. Under it, AI companies can voluntarily submit their frontier models for risk assessment 30 days before release. The fact that one of the first models under this regime has now been hit with an abrupt access stop is raising alarm bells across the industry. Industry associations and think tanks are calling it an ad hoc licensing regime under which a small group in the White House decides if a system is allowed to stay on the market.
The security community has also voiced criticism. Dozens of experts are calling for the restrictions to be lifted in an open letter. They point out that AI has long been used at scale for bug-hunting and exploit generation, that models like Mythos may be powerful in this space, but are not unique, and that the ban primarily affects teams responsible for securing production systems and legacy code. They argue that cybercriminals can simply fall back on open-weight models or older techniques. The US government is, in effect, taking an important tool away from its own companies while potential attackers carry on regardless.
Internationally, the case has reignited the debate about independent AI models. In Europe, politicians are waking up to the fact that countries are making themselves dependent – and therefore vulnerable – by relying on a handful of US providers. The abrupt ban of Fable 5 and Mythos 5 is widely seen as a warning sign that, when it comes to the crunch, Washington will act in its own security interests – without regard for allies.
Europe has so far struggled to develop its own AI models and lacks a comparable server infrastructure. France’s Mistral is relatively capable, but it can’t keep pace with Claude, ChatGPT or Gemini. Apertus, ETH’s open-source model, lags even further behind and wasn’t primarily designed as a chatbot.
Dispute could have a negative impact on the IPO
For Anthropic, the conflict also presents a financial problem; it’s planning to go public soon. Having a flagship model banned and a regulatory target on its back could put investors off. In its statement, the company stresses that it supports the government’s right to take action against genuinely unsafe models. But Anthropic is calling for a transparent, technically grounded process, arguing that the action taken doesn’t meet these criteria.
One thing is clear: this affair marks a turning point. For the first time, the US government has used export controls to restrict access not just to hardware, but to a specific AI model. Customers have experienced first-hand how an important tool can vanish overnight. From now on, some people will consider more carefully the political risks that come bundled with their choice of AI platform.
My fingerprint often changes so drastically that my MacBook doesn't recognise it anymore. The reason? If I'm not clinging to a monitor or camera, I'm probably clinging to a rockface by the tips of my fingers.
Interesting facts about products, behind-the-scenes looks at manufacturers and deep-dives on interesting people.
Show all
