Executives at the artificial intelligence start-up Anthropic received alarming news from the White House on Friday. They had less than 90 minutes, they were told, to take down their newest A.I. models over national security concerns.
Inside the company, employees’ group chats immediately lit up. Managers were instructed to prepare customers for a potential service disruption to the models, called Fable 5 and Mythos 5. But the messaging kept changing, with workers initially being told that the security problem was the ability of foreign companies to gain access to the systems, and later that a major vulnerability had been discovered in the models.
In employee chats, Anthropic engineers asked one another if the company’s plan to go public this year would be harmed by the White House directive. Many shared news reports that offered conflicting information about why the White House had ordered Anthropic to suspend access to Fable 5 and Mythos 5 for all foreign nationals.
“What are you telling your clients?” one employee asked in a chat viewed by The New York Times. Another said, “Does anyone know what to believe?” In another message, a worker said, “I don’t understand what the issue is.”
Six days later, Anthropic’s roughly 3,000 employees still have few answers. The San Francisco company is continuing to grapple with internal confusion as Dario Amodei, the chief executive, and some of his lieutenants meet with the Trump administration to try and resolve the situation. But after discussions on Monday and Tuesday, there was no breakthrough over ending the U.S. order to limit access to the company’s new A.I. models.
In a statement on Monday, Anthropic said it would continue meeting with government officials and pledged its “ongoing commitment to working alongside the administration.”
The dispute highlights how singular Anthropic has become in Washington. It was the second time in six months that the fast-growing A.I. start-up has become embroiled in a fight with the Trump administration over its powerful technologies, even as other A.I. companies offer similar models that have not received the same attention. And it has left Anthropic’s employees in what they described as a holding pattern, with some wondering if they were being picked on by President Trump.
“Are we being bullied based on bad vibes?” one employee asked in a chat viewed by The Times.
A White House official familiar with the discussions said the administration was continuing to engage with Anthropic to address its concerns.
Anthropic and the Trump administration first came to blows earlier this year over a $200 million Department of Defense contract for A.I. in classified systems. In a highly public argument, the company and the Pentagon disagreed over how the technology should be used in warfare.
That ultimately led Defense Secretary Pete Hegseth to label Anthropic a “supply chain risk” in February. The designation, which meant the company posed a risk to national security, had never been used against an American company. Anthropic has sued the government over the label.
The company’s relationship with the Trump administration changed in April when it announced Mythos, a new A.I. model. Mythos was so powerful at identifying security vulnerabilities in software, Anthropic said, that it could set off a cybersecurity “reckoning.” The start-up would hold back the model, it added, except to a select few organizations and companies.
At the White House, that prompted officials to begin discussing an executive order on A.I., which would include a voluntary process for companies to have new A.I. models reviewed before release. Anthropic took part in discussions around the order and spoke with the White House chief of staff, Susie Wiles, in a meeting some hoped would lead to a rapprochement.
Then this month, Anthropic released a straitjacketed version of Mythos, called Fable 5, which researchers believed was safe for widespread use and included additional guardrails. Anthropic provided the models for testing to the Commerce Department, as it had done for previous models, and was not told not to deploy them, people briefed on the events said. The results of the tests are not known.
Fable 5’s release prompted other companies to try out the model. One was Amazon, which has pledged to invest up to $33 billion in Anthropic and supplies some of the chips that power its A.I. models.
Amazon researchers wrote a preliminary paper on Fable 5 that pointed to a perceived security shortcoming in it, Anthropic representatives, U.S. officials and others said. Using that vulnerability, Amazon was able to convince Fable 5 to disclose flaws in specific bits of vulnerable software code. Amazon did not include tests of models from other A.I. firms, some of which are capable of generating the same information, cybersecurity experts who have seen or been briefed on the report said.
After Amazon shared the findings with Anthropic, the paper came up in a previously scheduled call between Amazon and administration officials, according to people briefed on the events. The officials asked Amazon to have a follow-up call with Scott Bessent, the Treasury secretary, to explain the vulnerabilities more fully. Amazon agreed.
An Amazon spokesman did not address specific questions about the sequence of conversations with the White House but said in a statement: “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.”
U.S. officials who reviewed the Amazon document said the results were “scary.” But cybersecurity experts said the ability to ask Fable to identify flaws could be more useful for cyberdefense than for malicious hackers. Everything the Amazon paper accused Anthropic’s technology of being able to do could also be done by OpenAI’s latest model, people familiar with the technology said.
Katie Moussouris, a prominent cybersecurity expert who reviewed Amazon’s work at the request of Anthropic, said in a blog post that the concern about Fable was not something to fix, because that would actually undermine its ability to help defenders secure their software code.
“Defenders need to be able to ask A.I. to fix the bugs in a file, explain why the fix matters and write tests that confirm the patch works,” Ms. Moussouris said. “That is not a guardrail bypass. It is the most valuable thing an A.I. model can do for defensive security.”
One government official said the problem with Fable 5 went beyond the Amazon paper and included unspecified national security concerns over which companies Anthropic chose to work with. That was not a point raised directly with Anthropic, three people with knowledge of the discussions said.
On Friday, the government called Anthropic and demanded that it pull the newest model off the market, disabling access for all clients. Within 15 minutes, executives were on the phone with U.S. officials, asking for the reasons behind the national security concerns, three people briefed on the call said.
But no further details were given. The start-up found out about the Amazon research paper on its own, the three people said, adding that Anthropic had approached the White House and asked if the paper was the cause of concern.
Some cybersecurity experts said Anthropic was being unfairly targeted by the Trump administration. By Tuesday afternoon, more than 150 of them had signed an open letter to Howard Lutnick, the commerce secretary, and Sean Cairncross, the national cyber director, calling for them to lift the restrictions on Anthropic’s models.
The list of signatories included cybersecurity luminaries credited with large contributions to internet security, as well as an A.I. expert at the chip giant Nvidia and a former top official at the National Security Agency who oversaw responsible A.I. use.
“Anthropic has built multiple protections into the Fable model to prevent its use for cyber offensive uses,” the letter said. “These protections were so aggressive as to be the source of humor in the cyber community on launch day.”
At Anthropic, some employees shared the letter with one another, seeing it as proof that they were being unfairly targeted. On company channels, workers also asked how their employer planned to move forward if the White House continued limiting their ability to release new A.I. models.
“At what point does this just feel like they don’t want us to exist?” one employee said in a chat on Tuesday.
