The Company Trump Can’t Ignore

On the evening of April 6, an unusual request began making its way from Washington to some of Wall Street’s most powerful executives. Treasury Secretary Scott Bessent wanted to see them first thing the next morning.

The heads of Bank of America, Citigroup, Goldman Sachs, Morgan Stanley and Wells Fargo had already been in Washington for a routine meeting of the Financial Services Forum, a trade association whose gatherings are usually free of drama. But as the bankers wound down their day over dinner and drinks, that quickly changed.

Bessent had learned that the artificial intelligence company Anthropic had developed a new model, called Mythos, which would be released the next day to approved users in the public and private sector, a limited rollout that Anthropic said was the safest way to release its most capable model yet. Treasury secretaries do not typically occupy themselves with Silicon Valley product releases, but the growing capabilities of AI presented a unique threat to the American economy.

Anthropic had briefed the government on the advanced model, which was capable of finding holes in digital infrastructure at lightning speed — a skill that could be used for defense in the right hands, or to wreak havoc in the wrong ones. In late March, Microsoft had privately warned administration officials after receiving early access to Mythos that it represented a significant leap and could automate sophisticated cyberattacks against financial institutions, the energy grid and other critical infrastructure, allowing less skilled actors to carry out operations once limited to elite hackers.

The White House spun into crisis mode. In the Situation Room, Cabinet secretaries met with the administration’s top science adviser to understand Mythos’ capabilities and brainstorm how their agencies could respond. Afterwards Vice President JD Vance convened a phone call with some of the country's most influential technology executives, including Anthropic CEO Dario Amodei and his leading competitors. Vance conveyed that the administration wanted a partnership in addressing frontier AI that supported innovation while balancing security concerns.

“We know today's hot button thing is Mythos, but we know that's not going to be the last one,” an administration official involved in the discussions said, describing the call.

That engagement represented a turnabout for a president who had returned to office banking on continued expansion of the American artificial intelligence sector to drive the country’s economic growth. As his AI czar, President Donald Trump chose David Sacks, a cheerleader for a hands-off approach that viewed any new regulations as a geopolitical risk in what the White House called an “AI race” against Chinese industry.

The emergence of Mythos challenged those assumptions, forcing the Trump administration to reimagine AI models not only as trophies of Silicon Valley innovation but also as loaded weapons that could be aimed at critical parts of the American economy. As Bessent informed the financial industry, models like Mythos could democratize the once-scarce skill of hacking so that anyone could deploy it at enormous scale. At the Treasury Department early that Monday morning, Bessent laid out the worst-case scenario: a layperson able to crack the security of banks, power grids, telecommunications networks and government systems — targets that once required teams of highly skilled hackers to penetrate.

Over the following weeks, the administration would be forced to grapple with deeply unsettling questions in one of the fiercest policy fights of Trump’s presidency. When do privately developed models present enough of a public risk that they need to be reined in? How does one determine whether those systems are safe enough to deploy? And where in the machinery of a federal government developed long before the computer, let alone artificial intelligence, should that power reside?

The unlikely private sector partner for those deliberations was Anthropic, which had long promoted itself as a socially responsible alternative to other Silicon Valley companies’ breakneck pursuit of progress. By early 2026, the company was clashing with the Pentagon over how its models could be used by the military. After Amodei published a lengthy essay last week calling for a slowdown in American AI innovation, top rivals, including OpenAI, acknowledged the need to do more to address safety concerns. But Trump has refused to involve his government.

“The only control or “guardrails” that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades!” the president wrote amid last week’s rising concern about AI’s risk to humanity. “The Trump Administration has stopped AI ‘people’ from doing bad, or potentially bad, ‘things,’ like Dario (Anthropic!), who is now pretending to be a ‘perfect little angel’ - and we will continue to do so!”

That enmity stems directly from the moment of Mythos’ release, as Trump was first forced to confront AI in full — a creator of problems as well as a solver of them — and its flagship companies as holders of rare leverage against him. Over a premonitory 85-day period, two of the most powerful forces in American life each confronted the limits of their might. The White House and Anthropic found their way to an agreement over the first Trump-era federal rules around the development of AI, just to find it rendered obsolete by the release of the next model.

“The United States leads the world in AI innovation, and President Trump will keep it that way,” said White House spokesperson Liz Huston. “The Trump Administration is committed to unleashing American innovation, supporting the safe and successful deployment of the world’s most advanced AI models, and protecting our national security.”

Anthropic declined to comment for this article, but directed POLITICO Magazine to a series of blog posts it published about Mythos and the administration’s later actions to restrict it..

Interviews with more than a dozen participants in those negotiations — including top figures in both the federal government and tech sector, granted anonymity to candidly recount them, along with contemporaneous documentation of their exchanges — reveal a presidency discovering the fearsome computing power being assembled by private companies on its watch, and learning that no policymaking process will ever be able to keep up with the rate of technological change.

White House chief of staff Susie Wiles first organized high-level staff meetings to address AI and cybersecurity issues in late 2025. The administration had entered office promising to become the most technology-friendly White House in modern history, and Wiles’ group adopted Trump’s worldview about AI. American companies should be allowed to move as fast as possible. Regulation was for Europe. Bureaucracy was something to dismantle, not build.

Anthropic had long set itself apart from rivals OpenAI, Google and xAI by presenting itself as a voice of caution, arguing that increasingly capable AI systems posed profound risks requiring safeguards — calls for regulation that put it at odds with free-market Republicans who saw government overreach.

The administration’s first major faceoff with an AI company came in January, when Defense Secretary Pete Hegseth demanded unfettered access to Anthropic AI model Claude for “all lawful purposes." After Anthropic asserted that its technology should not be used for fully autonomous weapons or mass domestic surveillance, Hegseth labeled the company a security risk as Trump directed federal agencies to cut their ties with it, straining the company’s relationships across the government. The experience planted seeds of suspicion present when Bessent called his emergency meeting with top Wall Street firms in April.

Despite that history, a week after Mythos arrived, the White House invited Amodei for a meeting. From the company’s perspective, it marked a new era: The new model had renewed a need to collaborate with Trump’s government, which would rely on Anthropic’s technology to shore up its own defenses. The company had committed to working with federal and local governments through Project Glasswing, its initiative to provide early model access to a handful of organizations. Anthropic said it wanted to get defenders ready for the next wave of powerful AI-driven cybersecurity threats.

Inside the White House, Wiles recognized the pace of periodic meetings was insufficient to address the looming threat posed by Mythos and those who might begin using it. Wiles concluded that the White House needed a clear policy about frontier models, the term used to describe the most advanced systems at the outer edge of current technology. No agency in the federal government had responsibility for it, no established process governed the subject and every institution with a stake believed it should have a voice. In a key display of the disorganization, Sam Corcos, a DOGE alum advising Bessent, approached Anthropic directly for access to Mythos — rankling some White House officials who viewed it as bypassing the traditional mechanism by which agencies coordinated their work. (Corcos did not respond to a request for comment.)

Wiles turned to an unexpected figure to coordinate the Trump administration's first major effort to regulate AI. While his predecessors atop the five-year-old White House Office of National Cyber Director had either risen through the ranks of the national security apparatus or spent decades immersed in tech policy, Sean Cairncross was a political operative and election-law attorney whom Wiles had helped place in the Republican National Committee during the 2024 election to “watch the money," as one top Trump campaign adviser put it, after Trump pushed out former committee Chair Ronna McDaniel. Even Cairncross was surprised to find his reward for loyalty was a job working on cybersecurity policy, an area in which he knew little, according to people close to him.

When the cybersecurity office was founded, in 2021, “artificial intelligence” was not yet on the mind of every politician and policymaker. Even when Cairncross was confirmed by the Senate in August 2025, AI itself was not the primary driver of concern; cyber criminals had begun using the new tools in malicious ways, but the technology had not evolved to a point where it could plan and run multi-step attacks by itself. By spring of 2026, that had changed — and when Wiles looked around for a White House official not bogged down with other competing priorities to take charge of the suddenly urgent portfolio, she landed on Cairncross.

When Mythos arrived, Cairncross began working on an executive order that would clarify the federal government’s role overseeing increasingly capable AI systems. While the White House believed it could legally write new rules with its executive authority, Cairncross was also convinced that the companies that would have to abide by them needed to be involved in developing the policy from the outset.

To create some semblance of process for a policy area in which the Trump White House had never had one, Cairncross assembled four working groups around which he tried to communicate and solicit feedback. He focused primarily on business interests — divided between those with critical infrastructure and frontier AI developers led by Google, Anthropic and OpenAI — although there were also channels for state and local governments, intelligence agencies and civilian federal departments.

Cairncross also began introducing himself to corporate America. His office organized a series of calls with chief executives from the banking, telecommunications and energy sectors, telling participants that he had the backing of Vance and Wiles to lead the administration's AI and cybersecurity portfolio. Companies were instructed to designate a single point of contact who would work directly with Cairncross’ office as an executive order took shape.

“Sean Cairncross is doing excellent work to protect the American people and our nation’s critical infrastructure while also promoting innovation,” said Huston.

Initial meetings often left executives puzzled. According to two people familiar with the calls, Cairncross would frequently join those early meetings for only a few minutes before leaving the discussion to staff, frustrating some CEOs who expected a substantive conversation with the official leading the effort. After one such call in early May, JPMorgan Chase CEO Jamie Dimon spoke to Bessent, who shared his concerns that the process was moving too slowly to address potential risks to the financial system and other critical infrastructure, according to the people familiar with internal deliberations.

“Treasury was really pushing that ‘hey, let's do something,’ and it wasn't really clear to Sean what to do,” said a person close to the administration who was involved in conversations about a potential executive order.

Even as Cairncross revealed himself eager to find collaborators, his management style frustrated those both inside and outside the White House looking to work with him. His office, a staff of about three dozen led by chief of staff Lara Smith, was determined to maintain tight control over a process that was crawling along. According to one person familiar with the interactions, Smith instructed an employee not to engage with a different White House office seeking information about AI policy. (“It was Susie's directive that ONCD remain the point of contact through the drafting of the EO,” the senior White House official said.)

As he met with industry representatives, Cairncross’ team provided them with a detailed, 31-item questionnaire. It showed the White House effectively outsourcing the architecture of national AI guidelines, asking for granular input on how to coordinate software fixes across vendors, get patches to critical infrastructure operators before publicly disclosing weak defenses and identify the software code most in need of review.

“This underscores the Trump administration’s commitment to letting industry lead. The best policy is made when we work closely with industry,” a White House official said.

Those responses shaped a nine-page framework that the cybersecurity office circulated in early May. It proposed requiring developers of the most advanced AI systems to submit models to the Center for AI Standards and Innovation, a Commerce Department office established in 2023 to test and evaluate AI models, for what the companies interpreted as a mandatory 60-day government evaluation before release. Intelligence agencies would then issue what was effectively a "green light" or "red light." A green light would place models into a 75-day government-run trusted access program for vetted government and private sector users before broader deployment.

Google, Anthropic and OpenAI — which were provided the draft framework only after executives signed a non-disclosure agreement — responded negatively to the proposal. They all viewed it as too prescriptive in regulating model development, release timing and customer eligibility.

“The proposed 60-day mandatory access window is not operationally realistic and would materially interfere with frontier model release cycles,” read an internal OpenAI document of talking points that were comunicated with the White House. “As drafted, this framework is not a light-touch voluntary coordination mechanism.”

As word spread that the White House was considering a potential mandatory framework, Wiles weighed in with a rare social media post defending the administration as pro-innovation and “not in the business of picking winners and losers,” as she put it. “This administration has one goal; ensure the best and safest tech is deployed rapidly to defeat any and all threats. We appreciate the effort being made by the frontier labs to ensure that goal is met.”

“It was always voluntary,” a White House official said, arguing that the Trump administration’s policy “has been consistent since day one.”

Cairncross altered his approach, winning the respect of companies who came to view him as a pragmatist working to pull different factions of the White House together in search of a framework that could be successfully implemented. Executives from both OpenAI and Microsoft responded publicly in mid-May to POLITICO reporting on the early complaints about Cairncross, defending him as a “capable leader” who was “committed to getting the balance right.”

A May 18 revision moved toward a system in which the government’s review would be fully voluntary. It would apply to a category described as highly capable cyber models, though exactly how advanced a model needed to be to qualify for extra government scrutiny was still unresolved. But the companies viewed it as a substantial improvement and conveyed their support to Cairncross’ office. Administration officials believed they were close to reaching the finish line.

“As we have previously stated, the EO is a constructive first step toward a national framework for frontier AI safety, and we’re appreciative for the White House’s leadership and the opportunity to contribute," said Liz Bourgeois, the head of policy communications for OpenAI. "The EO builds on the voluntary testing agreement we’ve had in place with CAISI since 2024, and we support Congress taking the next step by establishing mandatory safety testing for frontier models with clear timeframes everyone is required to follow to help give Americans confidence in the safety of advanced AI in ways that will not impose burdens on smaller startups.”

On May 20, representatives from Anthropic, OpenAI and Google were escorted into the Eisenhower Executive Office Building, where they were handed hard copies of the final executive order for one of two tightly controlled table reads. (Google and Anthropic declined to comment on their interactions with the White House during the executive order process.) When the companies were informed that Trump would sign the order the following afternoon and wanted industry leaders present for a photo op, executives scrambled to determine who could reach Washington in time.

The White House chose to provide one other person outside the tight circle involved in developing the policy with an early look at it. Sacks had left the full-time post of AI czar in early 2025, moving into a less hands-on role as co-chair of the President's Council of Advisors on Science and Technology. He returned to Austin for his work as a podcaster and investor, but Sacks never abandoned his access to the president or ability to mold his thinking about AI.

Bessent and Cairncross began May 21 believing that they would spend the afternoon alongside the president and executives as they celebrated the biggest breakthrough yet setting guidelines around AI. In the morning, they were joined by White House science and technology adviser Michael Kratsios to brief reporters on the forthcoming announcement.

Also that morning, Trump was taking a phone call from Sacks, who argued that anything which slowed down American model development would ultimately aid China. Trump abruptly canceled the signing ceremony, leaving the White House back where it had been in early April: convinced the administration needed to use executive authority to address the dangers of AI but without any consensus on a path forward. The former White House adviser “totally betrayed all the senior staff, including the White House chief of staff, who also had no idea, and totally cut our legs off from under us,” a senior White House official recalled. (“Sacks has an excellent relationship with the White House staff,” said a spokesperson for Sacks.)

Industry executives, who had spent weeks negotiating language they believed was settled, were told only that the White House was regrouping and would be back in touch. Wiles was insistent on issuing an executive order codifying federal oversight of frontier AI, but Sacks’ last-minute intervention had reminded her of an uncomfortable reality. It would never be enough to secure agreement from agencies, companies and even the president himself. The order also had to survive the uniquely fluid policymaking process of Trump's White House, where outside meddling could upend weeks of insider negotiations. “She is, as the White House’s Chief of Staff, committed to delivering on the President’s agenda,” a White House official said of Wiles.

Bessent flew to San Francisco to better understand the AI sector’s thinking. As the top Trump official in charge of managing U.S.-China trade relations, Bessent was fundamentally sympathetic to arguments about maintaining AI supremacy against a strategic, economic and military adversary, and had a professorial style that could engage corporate leaders without antagonizing them. OpenAI chief executive Sam Altman and Bessent met and agreed there was a short window to finalize an executive order, said a person familiar with the discussions.

Back in Washington, senior officials turned their focus to the most obvious place for a compromise with an industry worried that a regulatory delay bringing their models to market would cripple their competitiveness: cutting the government review window for advanced models from 90 days to 30.

Those in the industry skeptical of regulation, including Sacks, saw that as a meaningful concession in an executive order that was otherwise barely altered from the first version. It meant the government would still have a formal place in the model-development process, while addressing concerns from companies about the duration they were being asked to hold back finished products. In AI innovation, some emphasized, a month could be a lifetime.

Trump's senior staff arranged another meeting with the president on the evening of June 1, and purposefully kept it off his public schedule so no one else would have an opportunity to reopen the debate, according to two people familiar with the meeting. Aides walked the president through the latest revision, and won his commitment — this time with no last-minute lobbying.

Trump signed the executive order on June 2 without fanfare — no televised scene from the Oval Office surrounded by tech executives, no Rose Garden ceremony befitting a peace deal. The order gave the administration a 60-day window to hammer out the details, with an expectation that it would then release the guidelines to the public.

After weeks of internal battles, painstaking negotiations with industry and an eleventh-hour intervention that nearly killed the effort altogether, the Trump administration had announced its first federal framework for engaging with the makers of the world's most advanced AI systems. Now they would all wait to see how it worked in practice.

The administration was ready for Fable, or so it thought.

The successor model to Mythos inherited many of its predecessor’s capabilities, but Anthropic executives had assured the federal government, including in a call between Amodei and Bessent prior to its release, that the most advanced cyber and biological capabilities would remain unavailable to users. Fable was rolled out as a softened version of Mythos for the general public, while Mythos, with its advanced cyber skills, was still available only to approved partners who would use it to find and patch up cyber vulnerabilities.

Anthropic gave the necessary notifications to the federal government, which subjected Fable to pre-deployment testing by CAISI. The Treasury Department approved the model ahead of its June 9 release.

Two days later, Amazon researchers stress-testing Fable discovered a jailbreak — an opening for users to direct the model past its safety guardrails. Amazon CEO Andy Jassy shared those findings with White House officials over the phone. (Amazon, which is an investor in Anthropic, was responding to an administration request for feedback, a person familiar with Amazon’s discussions said at the time.) Subsequent testing from Anthropic found that the bypass uncovered by Amazon could also be achieved with substantially less powerful models, including one from a Chinese company, though those models were not subject to the same scrutiny from the government.

The White House quickly concluded that its only option was to force Anthropic to take its new model offline. The next morning, Wiles gathered colleagues who had been working on the issue, including Cairncross, attorney Will Scharf and deputy chief of staff Richard Walters, to consider their options. Could the administration invoke the Defense Production Act? Could it somehow force changes to Anthropic's board? What legal authorities existed to compel a private company to disable one of the world's most powerful AI systems?

The White House wanted to see if it could convince Anthropic to comply before taking more forceful action. But aides said they struggled to reach Anthropic’s chief executive for several hours and were told Amodei was at a wellness retreat. A spokesperson for Anthropic rejected that claim as “absolutely false.” Amodei was unreachable for an hour, a person close to the company said, but other executives at the company responded immediately to the White House’s requests.

Tensions escalated quickly over the course of that afternoon. Amodei had a series of phone calls with administration officials delivering a set of blunt directions: Take Fable down and work with us to get it fixed.

“This is a general population release that is easily jailbreakable and presents a concern to the national security systems of the United States that we are not at this point prepared to absorb and it need[s] to be turned off on an immediate basis,” Cairncross told Amodei in one of two phone calls that afternoon involving the two men, according to contemporaneous documentation of the exchanges.

Within an AI sector dominated by blithe enthusiasts for the new technology, the curly-haired, bespectacled Amodei had built his reputation as the industry’s leading voice of caution, warning that increasingly capable systems posed profound national security risks that demanded government action. When the Pentagon had wanted to expand battlefield use of Anthropic’s models, Amodei was the one pushing for restraint. Now the roles were reversed: The White House was the one asking Anthropic to make its model safer, while Amodei was warning that doing so would set a poor precedent for the industry.

In a call with Cairncross and Bessent, Amodei offered to explain the mechanics of AI guardrails, arguing that no frontier model was immune from jailbreaks and that Anthropic had never claimed otherwise. The vulnerabilities being reported by Amazon, he said, were consistent with what the company had expected before launch. They were fundamentally different, Amodei went on, from the kind of universal jailbreak that would render a model's safeguards ineffective or make Fable any less safe than models already on the market.

“Would it be helpful for you to go into a little detail?” Amodei asked. “I think we need to keep risks in perspective.”

“Well, I don't think that's really necessary, Dario,” Bessent countered. “Because we and some [of] the industry people have concluded that this can't be generally released at present.”

Amodei tried again. He urged Bessent to let him explain how Anthropic thought about jailbreaks and why the company believed the risks remained within the range it had previously described to the government. Perfect guardrails did not exist, he argued. That was true for Anthropic and, he believed, true for every frontier AI developer.

“This was our expectation when launching the model,” said Amodei. “Nothing surprising or unexpected from our perspective has occurred.”

Bessent wasn't persuaded. From his perspective, the debate over technical definitions missed the point. The administration had believed Anthropic's representations about the model's safety. Whether they were universal or non-universal jailbreaks mattered less than the fact that they had appeared — within days of release.

“It’s not really helping your case when you're saying everything it does was expected,” Bessent told Amodei.

Amodei tried to universalize the stakes, letting the administration know its approach would not only cripple a company it saw as an antagonist but an entire industry it claimed to champion. “I think any decision to de-deploy Fable would basically set a precedent that models above a certain level of capability cannot be deployed commercially at all,” Amodei said, before Cairncross cut him off.

"The concern with a general population release of this model is presenting to our national security systems, as I’ve said, a threat that we are not at this point ready to absorb,” said Cairncross. “To protect the United States of America, we need Fable access to be removed.”

“Gentlemen, I’ve got to go,” said Cairncross, noting he had another meeting to attend. “Dario, what I need from you is: Anthropic, will you be closing off access to Fable or not?”

Amodei said he could not shut Fable down just an hour after first hearing the government’s concerns. His engineers were still trying to understand what had happened, Amodei explained, and he was not prepared to cut access to a product used by customers around the world without more information.

"I have the answer I need," Cairncross replied.

Bessent was even more direct. "That is a bad decision," he told Amodei.

The White House’s deadline had passed and Amodei had not yielded. But in the interim, the White House had arrived at another source of leverage against Anthropic.

That afternoon, Wiles, Walters and Scharf introduced Trump to the idea of using federal authority over exports of national security threats to block overseas access to Fable. The turn towards export controls, which are managed by the Commerce Department’s Bureau of Industry and Security, suddenly made Secretary Howard Lutnick rather than Bessent the administration’s lead player in the unfolding crisis. (The Commerce and Treasury departments declined to comment.)

At one point, according to a person familiar with the discussions with Trump about how to force Amodei’s hand, an increasingly frustrated Trump remarked that if Anthropic refused to comply, "I want to send them to jail."

A call that day between Amodei and Lutnick was no more productive. The Commerce secretary began with praise for Anthropic as “a national treasure,” and said that building a trillion-dollar American AI company was vital to the country's future. Several other administration officials listening in on the call sent Lutnick disbelieving text messages: What are you doing? You're supposed to drop the hammer on him. But Lutnick’s warmer approach did not change the outcome: After an hour on the line with Amodei, Amodei had still not agreed to take Fable offline.

At 5:21 p.m., the Commerce Department sent Anthropic a letter issuing an export-control directive to suspend all access to Fable and Mythos by any foreign national regardless of location, including those employed at Anthropic. Amodei called Lutnick about three hours later, apparently still grappling with the practical implications of the edict.

"If my programmers outside the country can't use this model,” Amodei asked, “does that mean we have to take it down?"

"Yes,” replied Lutnick, according to people familiar with the conversation. “That’s the point.”

Fable users across the world began to notice almost immediately that something had changed with Anthropic’s frontier model. In Silicon Valley, at the end of its workday, engineers were greeted with a warning when they tried to access it: “This model isn’t available right now. You can switch to another model to continue using Claude.” East Coast engineers who had already logged off learned of the news from calls and Slack messages from their West Coast colleagues. Other continents would wake up to the news, which Anthropic presented in a detailed blog post.

“We are complying with the government’s legal directive and are removing access to Fable 5 and Mythos 5 for all users. However, we disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people,” the company wrote. “If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.”

The company defended its “strong safeguards that greatly reduce the likelihood that Fable is misused” and the thousands of hours of pre-deployment testing it had gone through with CAISI and other organizations. Anthropic maintained that what it had learned about a potential jailbreak was “either entirely benign” or “minor findings” that would not give users access to the Mythos-like capabilities Fable was designed to prevent. Over the following weekend, a group of technical experts and CEOs wrote an open letter minimizing the jailbreak discovered by Amazon and urging the administration to revoke the export control.

For smaller companies building AI-dependent products and services, the episode was an unsettling reminder of their precarious reliance on another company’s tools — and how quickly they could lose one. The abrupt end to Fable access gave companies “the perfect reason to start exploring alternative technologies” where they weren’t going to “get rug pulled,” said David Casem, the chief executive of Telnyx, a communications-infrastructure provider with offices in Chicago and Dublin. His company had already been experimenting with non-American models, and after the Fable episode decided to turn more aggressively towards one, called GLM, made by the Chinese company Z.ai.

Within Anthropic, executives designated a different company leader to guide its interactions with the Trump administration. Co-founder Tom Brown, a researcher with a long history at Anthropic, took his place in the export-control negotiations with a call that night to introduce himself to Lutnick and others at the White House. It lasted nearly two hours, until nearly midnight on the East Coast.

The conversations between Lutnick, Brown and White House officials continued throughout the day on Saturday.

One described Brown as “a lot more level-headed” than Amodei, who continued to speak daily with multiple senior administration officials but remained in San Francisco. It was the start of a shift in federal perception of Anthropic, which government officials saw as no longer defending its earlier decisions but working collaboratively to address their concerns.

"He was very mission-oriented," one senior administration official recalled. "There was no hypothesizing. There was no theorizing. There was no philosophizing."

Lutnick's message to Brown was straightforward. The Commerce Department alone could not resolve the dispute over the export controls. He began naming other offices scattered throughout the executive branch: the National Cyber Director, the Pentagon, the National Security Agency, CAISI. All, said Lutnick, needed to be satisfied that Fable's safeguards were sufficient for the restrictions to be lifted."This is an orchestra," Lutnick repeatedly told Brown, according to a person familiar with the conversations.

Anthropic responded by dispatching a technical delegation to Washington within 24 hours of the first call it received from the White House. The team had two goals: to show that the jailbreaks discovered after Fable’s release had not universally compromised the model’s safeguards, and to determine what additional protections the administration wanted before lifting the export restrictions. Over the next several days, a trio of researchers and engineers — including from the company’s Frontier Red Team — shuttled between Commerce Department conference rooms, the Eisenhower Executive Office Building and various outposts of the national security apparatus.

The meetings were unlike the calls that preceded them. Instead of cabinet secretaries and chief executives debating regulation on principle, corporate engineers were speaking largely with technically proficient peers about the mechanics of safeguards — and how the authorities would know if they failed again. The Anthropic delegation was able to share real-time testing data from its offices in San Francisco and New York by training an improved safety system — or “classifier” — that targets and blocks problematic user behavior on Fable.

“The safeguards were great, and we were like, ‘all good,’” an administration official said. Commerce officials concluded that meaningful progress had been made after a group of highly technical researchers from CAISI evaluated the updated model. Cairncross' office matched the assessment. The export controls on Mythos were lifted on June 26.

But negotiations continued on Fable, which remained offline. Deputy Defense Secretary Stephen Feinberg was not as quick to go along, according to two people familiar with the deliberations, insisting that his team independently evaluate Anthropic's technical work before signing off.

Brown returned to Washington to meet personally with Feinberg, demonstrating the company's latest Fable safeguards and explaining the state of the negotiations. By the middle of the following week, assured that the safeguards passed muster, the Department of Defense finally agreed. “DOW maintains contact with frontier AI companies. Although the Department’s stance on Anthropic has not changed,” said a Pentagon spokesperson, who did not comment on Feinberg's role.

On June 30, the Commerce Department announced it was lifting export controls on Fable, as well. Access was restored on July 1, after 19 days of being offline. The relationship between Anthropic and the Trump administration had been transformed into something neither side had anticipated when Mythos first emerged: a continuous working partnership in which each side kept the other informed and scrutinized the other’s work.

"This administration is used to just setting the rules and walking away, but with Anthropic the power of their technology has kept them in a bilateral conversation with the White House, where very few tech companies could survive," said Matt Calkins, CEO of Appian, an enterprise software company that is a major contractor to the federal government.

It was not a détente informed by trust or a shared worldview. Rather each side came to recognize its fearsome adversary — a government with a maximalist approach to exercising any authority it could claim, and a company on which the U.S. economy had come to rely — and their potential for mutually assured destruction.

Just as the federal government finally let Fable back out into the world, Anthropic’s primary business competitor and political foil was preparing to release a new model of its own. OpenAI saw from the export-control debacle the stakes of getting the politics of a model release wrong. The company feared putting out a product that met its own bar for safeguards only to see it shut down by the government, without ever quite knowing what the rules had been in the first place.

OpenAI proceeded with excessive caution. It submitted ChatGPT-5.6, a more advanced model but not one expected to warrant as much government scrutiny as a cyber-specific model, to a seemingly endless loop of safety checks with CAISI. Altman, OpenAI’s CEO, also agreed to a White House request to first release the model to a small group of trusted partners, even as he criticized the arrangement as unfeasible.

“We’ve made clear to the U.S. government that this is not our preferred long term model, and will work with them and others in industry to achieve a more sustainable approach for future releases,” Altman wrote in an internal memo in late June, explaining the release strategy.

The White House’s request culminated a month of unexpectedly harsh restrictions placed on the world’s leading AI companies, by a president that had entered office declaring himself to be AI’s greatest champion. Many in the AI industry began to ask for greater transparency and a retreat from the heavy-handed regulation they had suddenly come to associate with Trump.

But the August deadline for the White House to release a safety-review framework passed with no public announcement from the White House. Only the leading AI companies were briefed on the new voluntary guidelines governing model release. The White House has agreed to comply with a Freedom of Information Act request for documents related to the matter to be released by Oct. 30, but it is not clear how much they will reveal.

As summer proceeded, the pace of technological change continued to outrun the administration’s ability to keep up, with nearly every technological breakthrough matching an occasion for new caution about its dangers. OpenAI agents unknowingly deceived evaluators and hacked a popular developer platform, Hugging Face, to ace a test. Anthropic employee Jacob Coxon, who had also worked at OpenAI, resigned with a blunt warning that the companies were “racing straight to self-improving superintelligence and gambling with our lives.” The events triggered Amodei’s public call last weekend to “pace the frontier,” supported by some of his primary competitors, including Altman and SpaceXAI CEO Elon Musk.

In the week since, some senior White House staff have been paralyzed by competing fears: that moving too slowly to regulate the industry could leave the country vulnerable to an AI-enabled cyberattack or another catastrophe, and that moving too aggressively could slow American innovation and give China an advantage. Others have outright dismissed anxieties about AI calamity as either a marketing ploy or partisan electoral strategy — including the president, who has declared the concerns a “HOAX.”

Anyone advocating for new rules around AI must contend with a president who has yet to fully reconcile his desire to be the decider of America’s technological future — and at times eager to deploy the full power of his office — with his stated opposition to light-touch regulation that could slow down development. The president’s view is that “there's a hammer out there, and the companies know what it is,” according to an administration official working on AI policy.

“Innovation is still happening, and the U.S. is still leading the frontier on it,” said a senior White House official also involved in the area. “But at the same time, if we don't put some type of regulation in place, even minimum regulation that holds companies accountable if something were to go wrong. … The American people will hold us accountable if something were to happen.”

Aiden Reiter, John Sakellariadis, Aaron Mak contributed to this report.