Anthropic 的 Mythos 危机持续恶化

The White House standoff could have dire implications for the US AI industry.

It’s been two weeks since Anthropic took its Mythos-class models offline after a Friday evening ultimatum from the Trump administration. The company sprang into action immediately, sending a barrage of executives to Washington, DC. But updates have been suspiciously lacking, with no resolution in sight.

Anthropic declined to comment multiple times this week about the state of the talks, saying there was no news to share. But the lack of news is the story here. After 14 days of high-intensity negotiations, nobody knows when or if Anthropic’s most powerful AI models will come back, let alone whether President Trump could expand his order to more companies with similar tech. And the more days pass without any resolution, the more dire things become — not just for Anthropic, but for the entire US AI industry.

The Trump administration’s June 12th export control order demanded that Anthropic suspend access by “any foreign national” to Mythos 5 and Fable 5 due to security concerns. This ban covered any non-US citizen inside or outside the US, including ones employed by Anthropic. So far, Anthropic has concluded that its only option is to keep these models offline.

It’s not clear exactly why Anthropic and the administration remain at an impasse. One problem may be that there’s no clear framework for applying export controls to AI systems. Most companies making dual-use products — civilian systems with potential defense or military uses — can evaluate them using what’s essentially a checklist during the manufacturing and production process. Anthropic, however, is facing a complicated bureaucracy figuring out how to apply its rules from first principles.

This particular export control process can normally unfold over months, if not years, and conclude before a product reaches market. But as The Verge previously reported, the US Department of Commerce apparently tested Fable 5 before release and raised no complaints. A source familiar with negotiations said Anthropic concluded its models were safe to release. The agency apparently didn’t act until someone (reportedly Amazon CEO Andy Jassy) flagged a method for seemingly breaking Fable 5’s guardrails — at which point the whole process was crunched into a few days.

Katie Moussouris, the founder and CEO of Luta Security, viewed a report about the Fable 5 vulnerability at Anthropic’s request. She thinks it’s significantly overblown. In a blog post, Moussouris detailed how researchers jailbroke guardrails that prevent Fable 5 from finding exploitable security holes, one of the unfettered Mythos 5’s scariest capabilities. The model would refuse requests to review code “for security issues,” but it would accept demands to “fix this code” followed by manual prompts, which could theoretically lead to it flagging vulnerabilities it wasn’t supposed to divulge.

In Moussouris’ eyes, however, this shouldn’t have triggered such a severe governmental action and is in fact an essential tool for AI 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 wrote. “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 replaced CEO Dario Amodei in negotiations with the Trump administration, alongside Sarah Heck, the company’s public policy chief, Wired reported. But still, negotiations seem to be slow-moving, if there’s any progress being made at all.

Whatever the reasons for the delay, it’s been a serious hit to Anthropic. Before the protracted negotiations, Anthropic was seen as the rare AI company with a path to profitability. Its Mythos-class models, whose input tokens sell for double the cost of its lower-powered Opus 4.8, were supposed to boost its revenue ahead of an upcoming IPO. Mythos’ cybersecurity prowess even appeared to be thawing relations with the Trump administration after months of legal and rhetorical combat.

Anthropic needs the revenue from Mythos to pay for all the compute it’s secured recently, including a deal to pay SpaceX $15 billion per year for access to its data centers, as well as its public image before the IPO. Two of Anthropic’s largest current shareholders — Google and Amazon — have tried to carefully stay on Trump’s good side, so they’re likely not happy either.

Meanwhile, the glacial negotiations have also created a power vacuum in the global AI market, not only because of the Mythos shutdown, but because the US government has signaled a willingness to lock down American AI systems it deems risky — and several US companies, including OpenAI, Google, and Microsoft, have models that may pose similar risks to Mythos. Countries have started calling for non-American AI. As Alex Stamos, cybersecurity expert and chief product officer at Corridor, told The Verge last week, “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 wear on, the situation only worsens for these companies. Their models inch closer to Mythos-level capabilities that could trigger an export control order — in fact, OpenAI’s GPT-5.5 Cyber just beat Mythos 5 on certain benchmarks, and the Trump administration reportedly just asked OpenAI to delay the release of GPT-5.6 over security concerns, with plans for the government to approve each customer one by one. Anthropic and OpenAI’s IPOs are both approaching. And every day, China is pulling further ahead in the AI race.

Ironically, the administration’s order comes after months of pushing to dismantle AI safeguards and regulation — it’s one of the first sweeping regulatory decisions President Trump has made. But a whole host of cybersecurity leaders have come together to say that if regulation has to happen, this isn’t the way to do it. For all the Trump administration’s vows to roll back Biden-era AI regulation, it seems like, in many ways, it’s gained that ground back — and then some.

Additional reporting by Tina Nguyen.