Anthropic's Mythos Recall and the White House's Missing AI Safety Playbook

Justin Hendrix is editor of Tech Policy Press. This piece is co-published with Just Security.

On June 10, Anthropic founder and CEO Dario Amodei published a blog post titled, "Policy on the AI Exponential.” He wrote that in recent months “the evidence of AI’s incredible power, as well as its risks, has become undeniable.” He pointed to his company’s product, Mythos, as the "emblematic example” of the threat that frontier models pose to national security. “We need to activate a slow and rickety policy apparatus to deal with risks and opportunities that are going to compound surprisingly quickly from here,” Amodei wrote. He called for “mandatory testing by a qualified third party” of model risks in four areas: (1) cybersecurity, (2) biological weapons, (3) loss of control of AI systems, and (3) “automated R&D that could accelerate these other risks.”

Just two days later, on June 12, Anthropic announced it received a letter from the US Department of Commerce notifying the company that the government had “issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.” To satisfy the government’s demand, the company said it had disabled access to the two models for all its customers.

According to the Wall Street Journal, concerns raised by Amazon CEO Andy Jassy to the White House instigated the Friday letter. The government, Anthropic said, “believes it has become aware of a method of bypassing, or ‘jailbreaking’ Fable 5,” though the company said the risks were no different than with other models already on the market as it disputed the notion “that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.”

The Journal dubbed the episode “one of the most powerful examples yet of US government intervention in the AI race.” Former Trump administration AI advisor Dean Ball called the move “baffling,” pointing to the apparent inconsistency between the administration’s willingness to loosen export controls on advanced chips to China and the blanket restriction on access to Anthropic’s advanced models for all foreign governments, including close allies, many of whom were already relying on Mythos to identify cybersecurity vulnerabilities in critical systems.

Given these events, observers at the intersection of AI policy and security matters may feel a sense of whiplash. That’s particularly so given this latest dispute comes on the heels of recent policy announcements by the Trump administration that many regarded as a change in tune from the aggressively laissez-faire rhetoric officials advanced a year ago.

Regardless, the immaturity of the federal government’s approach to AI is now on full display. It is likely to have significant and reverberating implications for ongoing debates – and efforts to address – issues at the intersection of AI governance, technological “sovereignty,” and the relationship between US state and corporate power that are playing out in Washington and in capitals all over the world. It could be fairly argued that these ad hoc and wild swings on the part of the US administration is no way to exercise state power, and is further evidence of the need for a regulatory system that provides a more stable equilibrium for stakeholders to operate.

From ‘laissez-faire’ to lockdown

On day four of his second term in office, President Donald Trump issued an executive order titled, “Removing Barriers to American Leadership in Artificial Intelligence.” It rescinded Biden era policy, including a 2023 executive order on “Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence,” which the Trump order characterized as an obstacle to “global leadership in artificial intelligence.” The order effectively reset the federal government’s AI policy process. In doing so, it discarded Biden-era tools that were designed to surface the types of risks the Trump administration is reacting to post-Mythos, including a scheme for the Department of Commerce to collect red-team results for dual use foundation models.

What followed were a series of actions and orders that, together, revealed a more complicated picture of the administration’s AI policy. As Alondra Nelson argued in Science earlier this year, it became clear that the administration's approach involved "not the absence of AI regulation but its rearrangement," operating through industrial policy, trade restrictions, immigration controls, and strategic preemption rather than only traditional regulatory processes. And yet it is often difficult to locate a coherent strategy within that rearrangement. The same administration that sought to strip away oversight in January 2025 is now, in June 2026, fumbling toward using ever more drastic levers of state power to exert control over the AI industry, and Anthropic in particular.

After dithering on whether to release it, on June 2, the administration published an executive order "Promoting Advanced Artificial Intelligence Innovation and Security." The order mandates actions to integrate AI into national security infrastructure and to set up a framework for “voluntary” federal oversight. To some observers, the order was a step in the right direction, even if—as Mariana Olaizola Rosenblat argued in Just Security—it “drew the right map” but “stopped at the trailhead” by declining to make the framework binding. Council on Foreign Relations analysts argued the executive order reflected “an administration trying to sustain its deregulatory, innovation-first posture while confronting the novel cyber risks posed by powerful new tools like Anthropic’s Claude Mythos Preview.”

Three days later, the administration released what it called a “historic directive” on AI and national security. The National Security Presidential Memorandum (NSPM-11) sets up an aggressive timetable for the government to issue new policies, develop new strategies, and update complicated documents such as the Department of Defense Directive 3000.09 on autonomy in weapons systems. NSPM-11 contains a provision asserting that “the national security enterprise shall ensure, through contractual clauses or other means, that no commercial entity or adversary possesses the capability to prevent use of, disable or degrade, or materially modify without Federal Government knowledge and approval, an AI system that our men and women depend on for their missions.”

Amos Toh, a senior counsel in the Brennan Center's Liberty and National Security Program, told me the NSPM-11 presents as a “study in conflicting impulses,” including the “unresolved tension between the desire to onboard as much commercial, proprietary AI technology as possible despite fear over any constraints or other choices contractors might bake into their products.” But now the government has compelled Anthropic to do to allied governments precisely what the memo insists no vendor should be able to do to the United States: cut off access to a model that is important, if not vital, to their security.

Some point out that Anthropic’s Amodei has played a substantial role in raising the temperature in Washington over AI threats. Former Meta AI head of research Yann Lecun posted on Facebook that Amodei’s “ridiculous fear mongering” is to blame for these current events. “One reaps what one sows,” he wrote. Adam Thierer, a policy analyst at the R Street Institute, acknowledged on X that Anthropic has “relentlessly raised the regulatory temperature in Washington by inviting far-reaching controls of frontier models.” But, he wrote, the administration’s action on Friday nevertheless represents “a significant escalation in the politicization of AI and centralization of control over advanced computation in this country.”

Can allies rely on American AI?

This set of circumstances is likely to raise alarms in foreign capitals about the reliability of American AI for critical applications. Putting no formal, binding policy framework with some kind of legible process in place while being willing to shut down a model after deployment suggests the United States will operate ad hoc, potentially with access to advanced models revoked on a case by case basis. To make matters worse, under this administration in particular, the climate is one of a cloud of suspicion that senior officials are picking favorites based on personal and political factors.

Tensions over “tech sovereignty” were already at a boiling point, with policymakers from Canada to Europe taking actions to reduce reliance on American technology and build alternatives. The issue is likely to be an undercurrent this week at the G7 Summit in France, as my colleague Mark Scott points out in Tech Policy Press. Allies “are fearful their countries or regions will become increasingly dependent on already-dominant US tech giants that have shown a willingness to bend to the White House’s political objectives over those of long-time allies,” he writes.

UK and European political leaders are already expressing precisely these sorts of fears over the move to restrict access to Fable and Mythos. Bruno Retailleau, a former French interior minister and now a presidential candidate, wrote on X that “a nation that depends on others for its technology is a nation that can be unplugged overnight.”

In his June 2 blog post, Amodei called for activating “a slow and rickety policy apparatus to deal with risks and opportunities that are going to compound surprisingly quickly from here.” This month’s whirlwind events suggest “rickety” may be generous. The White House’s improvisation on what appears to be the first real test of what to do about a frontier model’s risks may ultimately imperil its stated goal to ensure American leadership in AI.