Anthropic’s Mythos Situation Continues to Deteriorate
It’s been two weeks since Anthropic took its Mythos-class models offline following a Friday ultimatum from the Trump administration. The company quickly mobilized, dispatching a team of executives to Washington, DC. However, updates have been notably scarce, leaving many in the dark about the situation.
Anthropic has repeatedly declined to comment on the ongoing discussions, stating there’s no news to share. Yet, the silence speaks volumes. After 14 days of intense negotiations, there’s still no clarity on when, or if, Anthropic’s advanced AI models will return, nor whether the Trump administration might extend its restrictions to other companies with similar technologies. As time drags on without a resolution, the stakes grow higher—not just for Anthropic, but for the entire US AI sector.
The Trump administration’s export control order from June 12 mandated that Anthropic restrict access to Mythos 5 and Fable 5 for “any foreign national” due to security concerns. This ban applies to all non-US citizens, regardless of their location, including those employed by Anthropic. Consequently, Anthropic has determined that keeping these models offline is its only viable option.
The reasons behind the stalemate between Anthropic and the administration remain unclear. One potential issue is the absence of a clear framework for applying export controls to AI systems. Companies producing dual-use products—civilian technologies with potential military applications—typically have a straightforward checklist to follow during manufacturing. In contrast, Anthropic is grappling with a complex bureaucracy as it tries to navigate these rules from the ground up.
Normally, the export control process can take months or even years, often concluding before a product hits the market. However, reports indicate that the US Department of Commerce tested Fable 5 prior to its release and raised no concerns. A source familiar with the negotiations mentioned that Anthropic believed its models were safe to launch. The situation escalated only after someone, reportedly Amazon CEO Andy Jassy, pointed out a method that seemed to bypass Fable 5’s safeguards, prompting a rapid response from the government.
Katie Moussouris, founder and CEO of Luta Security, reviewed a report on the Fable 5 vulnerability at Anthropic’s request and believes the concerns are exaggerated. In a blog post, she explained how researchers managed to bypass guardrails that prevent Fable 5 from identifying exploitable security vulnerabilities—one of the more alarming capabilities of the unrestricted Mythos 5. While the model would refuse to review code for security issues, it would accept requests to “fix this code” followed by specific prompts, potentially leading it to reveal vulnerabilities it shouldn’t disclose.
Moussouris argues that this capability shouldn’t have triggered such a drastic governmental response and is, in fact, a crucial tool for AI in coding. “Defenders need to be able to ask AI to fix the bugs in a file, explain why the fix matters, and write tests that confirm the patch works,” she stated. “That is not a guardrail bypass. It is the most valuable thing an AI model can do for defensive security: executing the find, fix, and test loop defenders run every day.”
In the past week, Anthropic cofounder Tom Brown has taken over negotiations with the Trump administration from CEO Dario Amodei, alongside Sarah Heck, the company’s public policy chief. However, progress appears to be slow, if it’s happening at all.
Whatever the reasons for the delay, it has significantly impacted Anthropic. Before these drawn-out negotiations, the company was viewed as a rare AI firm on the path to profitability. Its Mythos-class models, which command a higher price per input token than its less powerful Opus 4.8, were expected to drive revenue ahead of an upcoming IPO. The cybersecurity capabilities of Mythos even seemed to be improving relations with the Trump administration after a period of legal and rhetorical conflict.
Anthropic relies on the revenue from Mythos to cover its recent computing expenses, including a $15 billion annual deal with SpaceX for access to its data centers, as well as to maintain its public image before the IPO. Two of Anthropic’s largest shareholders—Google and Amazon—have been careful to maintain a positive relationship with Trump, so they’re likely feeling the pressure as well.
Meanwhile, the slow negotiations have created a power vacuum in the global AI market, not only due to the Mythos shutdown but also because the US government has indicated a willingness to restrict American AI systems it considers risky. Several US companies, including OpenAI, Google, and Microsoft, have models that may present similar risks to Mythos. Other countries are beginning to call for non-American AI solutions. As cybersecurity expert Alex Stamos noted, “One of America’s champions is being kneecapped by the US government while we’re in a race with the Chinese. It’s just incredibly stupid.”
As the days pass, the situation worsens for these companies. Their models are inching closer to Mythos-level capabilities that could trigger an export control order—OpenAI’s GPT-5.5 Cyber has already surpassed Mythos 5 on certain benchmarks, and the Trump administration has reportedly asked OpenAI to delay the release of GPT-5.6 due to security concerns, with plans to approve each customer individually. Both Anthropic and OpenAI are approaching their IPOs, and with each passing day, China is advancing further in the AI race.
Ironically, the administration’s order comes after months of efforts to dismantle AI safeguards and regulations—marking one of the first significant regulatory actions taken by President Trump. However, a coalition of cybersecurity leaders has come together to argue that if regulation is necessary, this isn’t the right approach. Despite the Trump administration’s promises to roll back Biden-era AI regulations, it appears to have regained much of that ground—and then some.
