Amazon CEO의 미국 관료 면담이 Anthropic 모델 규제를 촉발

Shortly afterward, White House officials held a meeting to discuss how to respond and security researchers began testing Amazon’s claims. The group decided that the most direct way to address that risk was by preventing foreign governments, companies and individuals from accessing the tool, the people said. President Trump later signed off on the action for security reasons despite concerns about hindering innovation, a senior White House official said.

The administration had long felt that Anthropic, one of the leaders in America’s AI race, couldn’t be trusted to manage the security risks its new model presented. Friday’s call between some administration officials and Anthropic Chief Executive Dario Amodei reinforced that feeling, the people said. 

In response to the government’s move, Anthropic said it was shutting off access to Mythos and Fable to all users to make sure it complied, potentially hampering efforts by companies around the world to use the tools to identify software vulnerabilities. Many of the company’s researchers are foreign born, meaning the government’s rule effectively prevented them from working on the latest models, Anthropic said. 

The talks with Amazon—a big investor in Anthropic that supplies the AI company with chips for data centers, while deploying its best models to identify software vulnerabilities—are a sign of how America’s largest companies and governments are navigating the emerging technology’s novel capabilities. Friday’s rapid events show how quickly new discoveries and experimentation can affect government restrictions and, potentially, company fortunes.

“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. When they occur, we don’t share the details of these discussions,” an Amazon spokesman said. 

Anthropic has said that the vulnerabilities like those flagged by Amazon are relatively basic, that other publicly available models are capable of discovering them, and that they don’t represent a full so-called jailbreak, a point of view shared by some security researchers familiar with Amazon’s research. 

The startup said it has adequate safeguards in place and is known for giving priority to safety. It previously held off expanding access to Mythos at the direction of the White House.

The Information earlier reported that Jassy raised concerns with senior administration officials.

Combined with a recent executive order giving security officials more oversight of models, and discussions about the government’s potentially taking equity stakes in AI companies, the moves show a dramatic increase in the control the administration is willing to exert over the industry, some analysts said.

“We are talking about a significant escalation in the politicization of AI and centralization of control over advanced computation in this country,” said Adam Thierer, a senior fellow at the R Street Institute think tank focused on technology and innovation.

Shutting down the top models could hold back Anthropic as it prepares for an initial public offering as soon as this fall if users turn to other models. It could also benefit competitors such as OpenAI, which has a powerful cyber model of its own that it is gradually offering to customers and discussing with the Trump administration.

Founded in 2021 by Amodei and other OpenAI alumni who felt the ChatGPT maker wasn’t emphasizing AI safety, Anthropic has become a top model provider to businesses for the strength of its Claude Code tool.

The rapid-fire moves Friday reignited a long battle between Anthropic and the Trump administration, which remains distrustful of the startup’s ties to donors to liberal causes and its warnings about the dangers of AI. Anthropic previously hired several former Biden administration officials and Amodei has criticized Trump and his administration.

Anthropic and the Trump administration have been sparring over the use of the startup’s AI tools by the military, resulting in an unprecedented designation by the Pentagon that the company is a security risk. Anthropic is fighting the move in two separate lawsuits.

Kate Koren, a former official at the Commerce Department, now at the Center for Strategic and International Studies think tank, said she understands the security concerns but it looks like the White House’s dislike of Anthropic could have influenced the decision. The senior White House official said the move was about model safety and the Department of Defense wasn’t heavily involved.

Friday’s action comes after reports that a user had bypassed the safety guardrails of Anthropic’s Fable 5, adding to worries about its ability to facilitate cyberattacks.

An Amazon report on Fable shared with cyber experts showed that it was able to dig up security bugs in at least four software programs when the tool was prompted with a specific set of queries, according to people who viewed the paper. That is information that Fable would typically not provide, but it is still a long way from dangerous cybersecurity information, according to Andrew Morris, founder of the cybersecurity firm GreyNoise Intelligence.

Many other tools can already reveal this information, but Anthropic’s software is known for being able to convert that bug information into working “exploit” code. Such code could be used to break into networks, but there is no evidence that Amazon’s researchers were able to access that capability, which is protected by Fable’s security guardrails, Morris said.

Jassy’s calls to administration officials were viewed by some as a general warning that quickly escalated into a wide Commerce Department ban on foreign users accessing Mythos and Fable, the people said. National Cyber Director Sean Cairncross and Commerce Secretary Howard Lutnick were involved in the conversations. The Commerce Department is in charge of export controls on critical technology.

Anthropic began previewing Mythos to Amazon and other tech companies earlier this year, working with the Trump administration on a phased approach. The goal was to let critical users apply the model to find software vulnerabilities and address them before more of the public had access.

An April meeting that included Amodei, Bessent and White chief of staff Susie Wiles was seen as an olive branch and peacemaking opportunity, but also included a discussion about security