미국 정부, 안스로픽 AI 수출 통제 유지

Comment Loader Save Story Save this story Comment Loader Save Story Save this story Trump administration officials concluded talks with Anthropic on Monday without lifting export controls that were imposed last week on the company’s most advanced AI models in response to jailbreaking concerns, according to three people briefed on the matter. The administration continues to believe that there are ways to disable some of the guardrails on Anthropic’s Claude Fable 5, effectively allowing users to access the more powerful cybersecurity capabilities of the company’s Mythos model, the people said. Anthropic has said for days that the administration’s concerns are overblown, a position it reiterated in working group meetings held at the Commerce Department with government researchers from Center for AI Standards and Innovation (CAISI) and the Office of the National Cyber Director Sean Cairncross, one of the people said. The meetings were also attended by Commerce secretary Howard Lutnick, who dialed in by conference call from the G7 summit in Evian, France. Cairncross himself did not participate, the person said. On Anthropic’s side, cofounder and chief compute officer Tom Brown and head of external affairs Sarah Heck have been leading the discussions. Anthropic’s head of frontier red teaming, Logan Graham, and senior security researcher Nicholas Carlini flew to Washington, DC for the talks. “Both parties are working quickly to get this resolved,” an Anthropic spokesperson said in a statement to WIRED. A White House spokesperson declined to comment. It was not immediately clear how any next steps might play out. The Commerce Department expressed a willingness to find a way to bring Fable 5 back online for consumer use, but it would likely be contingent on Anthropic fully resolving the jailbreak concerns, the person said. Ringing the Alarm The emergency talks have come at a fraught political moment for Anthropic, which was already in a prolonged fight with the Pentagon over whether its AI models could be used for certain military applications. The Trump administration was first alerted to the jailbreak concerns last week. Amazon CEO Andy Jassy called Treasury secretary Scott Bessent directly about the alleged vulnerabilities, which played a role in spooking the administration, the people said. Jassy’s conversation with the Trump administration was first reported by The Information. Alarmed White House officials tasked the NSA to help review the vulnerabilities. The NSA responded that it believed it was indeed possible to strip away Fable 5’s guardrails, prompting the administration to impose restrictions on the model. Lutnick then spoke with Anthropic chief executive Dario Amodei on Friday, as the Commerce Department drew up its letter imposing export controls on Fable 5. Over the weekend, after Anthropic cut off access to the model for all users, Lutnick was on multiple calls with Brown and Heck, according to a person with knowledge of the events. It’s unclear why Amazon, one of the largest investors in Anthropic, rang the alarm on Fable 5. “As a leading cloud provider that serves a large number of private and public sector customers, it’s not uncommon for governments to seek our counsel on potential security risks,” an Amazon spokesperson tells WIRED. “When they occur, we don’t share the details of these discussions.” Security Disconnect At the core of the conversations between Anthropic and the administration is a disagreement over the severity of the Claude Fable 5 jailbreaking concerns. In a blog post on Friday, Anthropic implied that the administration’s characterizations of the potential risks are overblown. Some cybersecurity researchers reiterated this position to officials on Monday, sending an open letter arguing that the export control action taken against Anthropic was unjustified. “Anthropic’s Mythos-class models are quite good at finding flaws and weaponizing exploits. However, they are not uniquely good at these tasks, and many of the undersigned individuals regularly use other foundation and open-source models for security audits and red-teaming every day,” the open letter reads. “As a result, this action has taken the best models away from defenders, created market uncertainty, and risked America’s AI leadership without any real risk to justify it.” Jailbreaking works by prompting an AI model in specific ways to circumnavigate its safeguards. Because Fable 5 is a version of Mythos with certain cybersecurity, biology, and chemistry guardrails in place, getting around those protections would effectively give users a version of Mythos. Anthropic has itself raised significant concerns about allowing Mythos to be used by the general public; however, it said on Friday that Fable 5’s safeguards were strong enough to allow for a public release. Researchers who evaluated Amazon’s findings say that the issues identified did not fully nullify Fable 5’s safeguards. “It wasn’t a jailbreak per se,” says Katie Moussouris, founder and CEO of Luta Security, who published an analysis after reading the Amazon paper. Moussouris emphasizes that regardless of whether the US government has proof of a full Fable 5 jailbreak, restricting the model’s ability to access to certain topics is a stopgap at best. “Most of us [in security research] think guardrails are speed bumps and shouldn’t be treated like security boundaries for skilled adversaries,” Moussouris says. “They only serve to slow down the less skilled.” Investors in Anthropic have also been working over the weekend, trying to assess how the company’s latest spat with the White House affects its corporate future, says another person close to the company. Some investors believe the US government is singling out Anthropic, and a competitor may not have faced the same reaction if they released a model similar to Mythos, the person says. The White House’s export control directive also raises broader questions for other AI labs aiming to release AI models with Mythos-level capabilities, and how they can do so in compliance with the US government. It’s now expected that AI labs give the White House early access to advanced AI models , and that they be extremely proactive about keeping the US government informed on model launches, according to AI lab leaders who spoke with WIRED. “The events over the weekend … are informative for everyone that the [US] government would be willing to take these steps," says Aidan Gomez, CEO of Cohere, a smaller AI lab based in Canada that offers enterprise tools. “No one can be naive to that reality.”