WASHINGTON—Talks between Anthropic and Trump administration officials continued Monday without a deal to resolve the security concerns that pushed the White House to restrict access to the artificial-intelligence company’s latest model, increasing urgency on both sides to find a resolution.
Ami Vora, Daniela Amodei and Dario Amodei of Anthropic.
After the Trump administration said foreign governments, companies and many individuals were prohibited from accessing the model, called Fable 5, on Friday, Anthropic shut down access for all users to comply with the new rule. It did the same for Mythos 5, a more powerful model also covered by the restrictions.
Pulling down the models because of government intervention driven by concerns about the adequacy of safeguards was an unprecedented move for a leading U.S. AI company. The longer the models are offline, the more pressure builds on Anthropic and the White House—which earlier feuded over acceptable military use of AI—to find a compromise.
The latest talks about Fable on Monday included senior Anthropic technical staff and were led by the Commerce Department and National Cyber Director Sean Cairncross, people familiar with the discussions said.
“This follows numerous virtual meetings we have held with the government since the administration’s initial outreach. Both parties are working quickly to get this resolved,” an Anthropic spokesman said in the company’s first public comments since Friday. “This is part of our ongoing commitment to working alongside the administration toward our shared goal of protecting U.S. critical infrastructure and the U.S. lead in cyber defense.”
Led by Chief Executive Dario Amodei and his sister Daniela, the company’s president, Anthropic prides itself on responsible development of AI. It is one of the few tech companies that has challenged the Trump administration on AI policy. The company was recently valued at $965 billion and is headed toward an initial public offering, increasing the stakes of a delayed shutdown of its latest models.
Some administration officials have said that a resolution should include an acknowledgment on Anthropic’s part that its rollout of Fable and communication with the White House could have been improved, the people familiar with the talks said. Anthropic on Friday said the security vulnerabilities it was informed about seemed relatively simple and that it had worked with the government extensively before releasing it.
Monday’s talks included top Anthropic security researcher Nicholas Carlini; Logan Graham, who leads the company’s team that evaluates models for risks; and Dave Orr, the company’s head of safeguards.
The White House is under pressure to show it can responsibly oversee the rapidly developing AI industry after oscillating between a hands-off approach and a more security-driven posture in recent weeks. Monday’s talks included a Commerce Department unit in charge of evaluating AI models before they are released, the people said. The unit evaluated Fable before it was released. Commerce Secretary Howard Lutnick wasn’t involved but participated in weekend conversations, they said.
The latest saga started late last week when Amazon researchers found a way to get around Fable’s guardrails days after it was released. The safeguards were presented as a responsible way to release the model, a general-purpose version of Mythos, which is capable of carrying out potent cyberattacks and has worried White House and private sector leaders with its potential.
Amazon Chief Executive Andy Jassy told administration officials including Treasury Secretary Scott Bessent that the issue was serious and worth evaluating, according to people familiar with the discussions. The head of Amazon Web Services, the tech giant’s cloud-computing arm, Matt Garman, has been involved in some discussions with administration officials, they said.
Amazon is one of Anthropic’s biggest investors and partners, supplying the company with chips needed to run its models. The company is also a major partner to the government on cloud computing.
Write to Amrith Ramkumar at amrith.ramkumar@wsj.com
