US Imposes Export Controls on Anthropic's 'Mythos' Amid Sweeping AI Policy Overhaul

The United States federal government has initiated its most sweeping intervention into the artificial intelligence sector to date, combining targeted national security restrictions with a new blueprint for federal legislation. In early June 2026, the White House issued a major executive order and unprecedented export controls on specific AI models, while a bipartisan congressional coalition simultaneously released a 269-page regulatory framework. This dual approach marks a definitive shift from a purely laissez-faire posture to a targeted crackdown on models deemed to possess advanced cyber capabilities.[1][3][4]

The immediate catalyst for this policy overhaul was the development of two unreleased frontier models: Anthropic’s "Mythos" and OpenAI’s "5.5 Cyber." Security researchers and intelligence officials warned that these next-generation systems demonstrate a tenfold increase in capability and speed compared to existing models. Specifically, the government expressed concern that threat actors could use these models to autonomously identify and exploit critical software vulnerabilities at an unprecedented scale, prompting immediate federal intervention.[2]

On June 2, President Trump signed the "Promoting Advanced Artificial Intelligence Innovation and Security" Executive Order. The primary mechanism established for civilian AI oversight is a voluntary framework that requests developers to provide the government with access to "covered frontier models" for up to 30 days prior to their public release. This window is designed to allow federal agencies to assess potential national security and cyber defense applications before the models reach the broader commercial market.[1][2]

Legal analysts emphasize that the executive order explicitly avoids creating new federal law, mandatory licensing, or preclearance requirements for the broader AI industry. This reflects the administration's stated commitment to minimizing regulatory burdens that could slow American AI dominance, focusing its binding directives strictly on hardening government networks and critical infrastructure rather than policing commercial development.[2][5]

However, in a sharp departure from the voluntary framework applied to the broader industry, the administration utilized national security authorities to impose strict, binding export controls specifically on Anthropic's Mythos-class models. The government cited acute fears that the model could be weaponized for offensive cyber operations, effectively blocking its international distribution and restricting access to a tightly controlled group of partners.[3]

This targeted action immediately triggered a fierce legal and political backlash. The Electronic Frontier Foundation (EFF) and allied civil liberties organizations filed an amicus brief in a lawsuit against the Pentagon, arguing that the export controls violate the First Amendment. They contend that the restrictions are arbitrarily punitive, noting that other large language models with similar offensive cybersecurity capabilities are only subject to the administration's "light-touch" voluntary review process.[3]

Beyond model restrictions, a core, binding component of the June 2 order directs federal agencies, including the Department of Defense and the Treasury, to prioritize the cyber defense of National Security Systems. The directive mandates that agencies issue binding operational guidance to facilitate access to AI-enabled cybersecurity tools for federal, state, and local authorities, as well as critical infrastructure operators like hospitals and utilities.[2][5]

To operationalize this defense, the order mandates the creation of an AI cybersecurity clearinghouse. Led by the Secretary of the Treasury, the National Security Agency, and the Cybersecurity and Infrastructure Security Agency (CISA), this hub is tasked with coordinating vulnerability scanning, validating discovered flaws, and prioritizing the distribution of software patches in collaboration with private industry.[1]

The executive order also signals a shift in how the government will handle malicious AI use, directing the Department of Justice to prioritize the investigation and prosecution of AI-enabled cybercrime. By focusing on the criminal misuse of AI to gain unauthorized access to systems, the administration intends to rely on existing criminal statutes rather than creating new regulatory agencies to deter bad actors.[1][5]

Just two days after the executive order was signed, the legislative branch introduced its own comprehensive framework. On June 4, Representatives Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.) released a bipartisan discussion draft of the "Great American Artificial Intelligence Act of 2026." If passed, this 269-page bill would create the first comprehensive federal AI governance regime in the United States.[4]

The draft legislation proposes binding federal obligations specifically for "large frontier developers." The bill defines this category as companies with over $500 million in annual revenue that have trained a frontier model. Legal analysts note that this financial threshold would capture major industry players including OpenAI, Anthropic, Google, Meta, and xAI, subjecting them to stringent transparency, cybersecurity, and whistleblower protection requirements.[4]

A central tension driving this federal legislative push is the rapid proliferation of state-level AI regulations. With federal legislation previously stalled, states filled the vacuum; laws such as California's Transparency in Frontier AI Act (SB 53) and Colorado's revised AI Act took effect on January 1, 2026, creating a complex and fragmented compliance landscape for developers.[4][6]

To address this fragmentation, the bipartisan federal draft includes a powerful preemption clause. The provision is designed to override state laws that specifically regulate the development of AI models, instituting a three-year sunset on state authority. This mechanism is intended to force a unified national standard while explicitly preserving state laws of general applicability, such as consumer protection and anti-discrimination statutes.[4]

The federal preemption effort aligns with the administration's broader pushback against state regulators. In December 2025, the White House established an AI Litigation Task Force specifically to challenge state AI laws in court, threatening to pull federal funding from states with regulations deemed overly onerous. However, the executive order exempts child safety, data center infrastructure, and state procurement from preemption, leaving significant regulatory power in state hands.[6]

Significant uncertainty remains regarding how the June 2 executive order will be implemented in practice. The text of the order does not explicitly define key terms such as "advanced AI" or "covered frontier models." Consequently, the actual scope of the voluntary framework and the classified benchmarking process will depend entirely on subsequent agency rulemaking over the next 60 days.[1][5]

The fate of the bipartisan legislative effort is equally uncertain. While the Obernolte-Trahan draft represents the most serious congressional effort to date, it remains a discussion draft competing for attention in a crowded legislative calendar. The events of early June 2026 ultimately illustrate a bifurcated American strategy: aggressively shielding national security assets from specific, highly capable models, while attempting to protect the broader AI industry from binding domestic regulation through voluntary frameworks and federal preemption.[1][3][4][6]