For the makers and artists relying on AI tools, this is a stark reminder that your creative software can vanish overnight due to geopolitical friction
Anthropic has taken two of its newest AI models, Claude Fable 5 and Mythos 5, offline to comply with an export control directive received on Friday afternoon from the US government. The move cites national security concerns as the primary driver for disabling these systems.
This incident represents the latest escalation in tensions between the tech giant and the Trump administration. While the directive specifically targeted access for foreign nationals, whether residing inside or outside the United States, including foreign staff of Anthropic itself, the company has chosen to cut off access for all its customers to ensure full compliance.
The friction is not new. Earlier this year, the Department of Defence under Trump labelled Anthropic a “supply chain risk” after the firm attempted to draw strict boundaries on how the US military could deploy its technology. That designation effectively barred government agencies and their contractors from using Anthropic’s systems, prompting the company to file lawsuits against the administration.
The Mythos model and its intended purpose
On Tuesday, Anthropic launched Claude Fable 5, a variant of its Mythos AI model featuring safeguards designed to prevent answers regarding cybersecurity, biology, and chemistry. This public release followed a collaboration with the US government and followed a limited rollout of the Mythos Preview AI model in April.
The initial goal was to allow organisations to leverage powerful cybersecurity capabilities to strengthen their own defences, while mitigating fears that the technology could be exploited by malicious actors to build sophisticated hacking tools.
A vague order and a narrow vulnerability
In a blog post published on Friday, Anthropic stated it received the directive at 5:21pm ET. The letter, the company noted, failed to provide specific details regarding the national security concern.
Anthropic explained that it understood the government believed it had identified a method to bypass, or “jailbreak” Fable 5. The company reviewed a demonstration of this technique, which was used to identify a small number of previously known, minor vulnerabilities. According to Anthropic, these flaws appear relatively simple and can be discovered by other publicly available models without needing a bypass.
The company argued that it has implemented robust safeguards to minimise the likelihood of misuse. Furthermore, Anthropic claimed the jailbreak identified by the US government was narrow and would not have made an attacker meaningfully more dangerous than they would have been using another AI model.
“To date, the government has only given us verbal evidence of a potential narrow, non-universal jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws,” the company stated. “Our understanding is that one potential jailbreak was shared with the government.”
Spokespeople for the White House and the US Commerce Department did not immediately respond to requests for comment.
A clash of principles
Anthropic CEO Dario Amodei argued in a policy essay earlier this week that he and the company support a fair, structured, and transparent government process for blocking the release of unsafe AI models. However, in the Friday blog post, Anthropic contended that the current action does not adhere to those principles.
Key takeaways
- Anthropic has disabled access to Claude Fable 5 and Mythos 5 for all customers, overriding the specific targeting of foreign nationals in the government directive.
- The shutdown follows a dispute over “supply chain risk” status and a specific government finding of a narrow jailbreak technique involving codebase analysis.
- Anthropic claims the identified vulnerabilities are simple and publicly discoverable, arguing the government’s action lacks the transparency and structure they advocate for in AI safety.
Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.
