X-Risk Daily

On 2 June 2026, President Trump signed an executive order establishing a voluntary framework for pre-deployment evaluations of frontier AI models that pose catastrophic cyber risks to critical infrastructure. The order directs companies developing frontier models to share them with the government for testing and, if a model meets a classified threshold for cyber capabilities determined by the National Security Agency, the government will have exclusive access for up to 30 days before the model is released to other trusted partners—an apparent effort to secure vulnerable systems before attackers can exploit similar capabilities.

The policy marks a dramatic reversal for an administration that, just seventeen months earlier, revoked the Biden AI safety executive order and dismissed concerns about AI risk. The shift appears driven by the April 2026 debut of Claude Mythos Preview, Anthropic's frontier model that demonstrated unprecedented ability to identify and exploit software vulnerabilities. Following Anthropic's announcement, the Treasury Department and Federal Reserve convened emergency meetings with major bank CEOs, while the International Monetary Fund warned that such models posed serious financial stability risks. Anthropic has restricted Mythos access to approximately 50 organisations under Project Glasswing, though the programme expanded on the same day as Trump's order.

The executive order tasks multiple agencies—including Treasury, the National Security Agency, and the Cybersecurity and Infrastructure Security Agency—with developing within 60 days a classified benchmarking process to assess AI models' cyber capabilities and determine what constitutes a "covered frontier model." The White House framed the order as an attempt to shore up defences while avoiding mandatory licensing or burdensome regulation. The framework remains entirely voluntary, does not specify what actions should follow if a model proves unacceptably risky, and covers only cyber capabilities—not biological or other catastrophic risks.

The shift in tone has been striking. Figures who previously opposed AI safety measures, including former White House AI adviser David Sacks and Senator Ted Cruz, have now endorsed some form of oversight. Earlier drafts of the order reportedly proposed a 90-day government access window; the final 30-day window reflects compromise between national security and anti-regulation factions within the administration. The order also establishes an AI cybersecurity clearinghouse to coordinate vulnerability discovery and patching across government and industry, acknowledging that AI systems are now capable of finding vulnerabilities far faster than human defenders can address them.

The National Security Agency has deployed Anthropic's Mythos model for offensive cyber operations, with approximately half a dozen Anthropic engineers stationed inside the agency to customize and operate the system, according to a Financial Times report citing people familiar with the arrangement. The model could be used to infiltrate networks in adversary nations including China and Iran, sources told the publication.

The deployment marks a significant escalation in how frontier AI systems are being used in national security contexts, representing what Tech Times described as the most operationally significant known deployment of a frontier AI model for state-level offensive cyber work. The engineers are working as forward-deployed staff inside NSA facilities, responsible for adapting Mythos for specific operational needs, though it remains unclear whether they are involved in active operations. The arrangement occurs despite a federal ban on Anthropic technology following a February designation by the Defense Department branding the company a supply chain risk—the first such designation ever applied to an American firm.

The conflict between Anthropic and the Pentagon began in January during negotiations over a $200 million contract, when the Defense Department demanded Anthropic make its Claude models available for "all lawful purposes." Anthropic refused, insisting on restrictions against mass domestic surveillance and autonomous weapons development. The NSA deployment appears to have been exempted from the broader Pentagon restrictions, underscoring tensions within the U.S. government over how to balance AI capabilities with safety concerns. In April, Axios reported that Anthropic CEO Dario Amodei met with White House chief of staff Susie Wiles and Treasury Secretary Scott Bessent to discuss Mythos use within government, with both sides describing the meeting as productive.

The deployment coincides with Anthropic's expansion of Mythos access this week to 150 organizations across 15 countries, up from an initial release to approximately 40 trusted partners. Anthropic initially restricted access to the model, contending that its offensive cyber capabilities were too dangerous for wider release. The expansion came on the same day President Trump signed an executive order creating a voluntary framework for government vetting of frontier AI models before public release, a move that followed Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell convening an urgent meeting with Wall Street CEOs to warn about risks posed by Mythos, according to PBS.

The involvement of Anthropic staff in supporting offensive military cyber operations raises fundamental questions about the boundaries between commercial AI development and national security applications, particularly given Anthropic's public positioning on AI safety and its ongoing legal battle with the Pentagon. The arrangement has drawn scrutiny over what it signals about the role of private AI companies in state-sponsored cyber operations, with Anthropic simultaneously fighting the Defense Department in court while embedding engineers at the NSA.

In an appearance on BBC Newsnight, Clark said the industry needs the ability to slow development, arguing "You want the option to be able to take your foot off the gas and put your foot on the brake."

Clark's warning centres on recursive self-improvement — the threshold at which AI systems can autonomously enhance their own capabilities. The concern gained urgency from internal Anthropic data showing that Claude currently operates on code "of which 80% the system wrote itself," with Clark telling the BBC that reaching 100% self-written code "is possible within two years." The statement accompanied a formal research agenda published the same day by The Anthropic Institute, warning that AI is already accelerating its own development and that recursive self-improvement "could come sooner than most institutions are prepared for."

In a separate interview with Axios published in early May, Clark offered a more specific timeline, predicting a 60 percent or greater chance that an AI model will fully train its successor by the end of 2028. The Anthropic Institute document warns explicitly of a possible "intelligence explosion" — a term historically confined to AI safety circles — in which systems suddenly improve at blinding speed once they achieve full autonomy over their own development cycle. The concept was first articulated by mathematician I. J. Good in 1966, who wrote that "an ultraintelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion.'"

Clark's call for "brake pedals" forms part of a broader Anthropic proposal published alongside his BBC interview. In a blog post co-authored with researcher Marina Favaro, the company urged major AI labs to consider a coordinated slowdown or temporary pause in frontier model development. The proposal echoes Cold War-era crisis infrastructure: Clark told Axios that rival nations dealing with existential technology during the Cold War "found ways to talk to each other," and that similar geopolitical coordination may be needed for AI. However, reporting notes that major AI companies have not publicly committed to pausing research, and that recent US regulatory action on AI did not mandate government safety testing.

Clark's statement is notable as a rare explicit warning from a frontier lab co-founder about loss-of-control scenarios. Anthropic has framed the disclosure as part of its commitment to transparency, with Clark telling Axios that the company's motivation "has always been: Tell the whole story" — whether discussing risks or potential abundance. The company said The Anthropic Institute will research mechanisms to verify any coordinated slowdown, though it remains unclear whether Anthropic or other labs are implementing emergency shutdown capabilities in practice.