This is a developing story. NCT has covered the export controls arc since the initial directive, including Commerce Secretary Lutnick’s criminal penalties letter, the legal challenge from constitutional scholars, JPMorgan blocking Claude in Hong Kong, and the Trump administration’s reversal framing Anthropic as a national security threat.

Three events on June 23 landed within hours of each other. Together, they outline a US AI export control regime that is simultaneously blocking American companies from doing business, cutting off the government’s own intelligence agencies from critical tools, and applying rules inconsistently based on which company built the model. The policy is now fighting for its own survival.

Legion v. United States: The First Commercial Lawsuit

A US company called Legion filed suit in federal court in Washington on Tuesday, according to Bloomberg. Legion builds AI tools for attorneys and legal operations. It is based in the United States. It employs Canadian nationals who work remotely from Canada.

When the Trump administration’s export controls took effect, Anthropic disabled Fable 5 and Mythos 5 access for foreign nationals, per the directive. Legion’s Canadian employees lost access to the models they use daily. The company’s lawsuit challenges whether the federal government can unilaterally cut off a US company’s access to AI tools based on where its employees sit.

This is the first commercial lawsuit directly challenging the export controls. Legal scholars have questioned the administration’s statutory authority to impose these restrictions (NCT covered that analysis on June 19). Legion’s filing moves the challenge from academic debate to active litigation. The suit tests a specific, practical question: can the government force US companies to choose between their workforce and their AI tools?

The implications extend well beyond Legion. Any company with distributed teams, which in 2026 means most technology companies, faces the same structural vulnerability. JPMorgan already blocked Claude access in Hong Kong as a preemptive compliance measure. Legion chose to fight instead.

The NSA Loses Mythos

The same day, The New York Times reported that the National Security Agency lost access to Anthropic’s Mythos model as a consequence of the export controls.

The irony is structural. NCT reported on June 23 that Mythos breached “almost all” NSA classified systems during an authorized red-team test, with six Anthropic engineers now embedded inside the agency as forward-deployed staff. The model is so capable that the NSA depends on it for cybersecurity operations. Congressional hearings have highlighted the agency’s reliance on advanced AI for network defense and threat detection.

The export controls, designed to prevent foreign adversaries from accessing frontier AI capabilities, have now prevented the US government’s own signals intelligence agency from accessing those same capabilities. The policy’s architects set out to protect national security. They cut off one of the agencies responsible for it.

This is not a theoretical problem. The Five Eyes joint statement issued on June 22 warned that frontier AI models capable of threatening governments and businesses are “months away.” If the intelligence community’s own assessment is that the threat is imminent, losing access to the most capable defensive tool available is a direct operational degradation.

OpenAI Gets a Pass

Axios published an analysis on June 23 documenting what many in the industry suspected: the administration is not enforcing consistent safety standards across frontier AI companies.

OpenAI released GPT-5.5-Cyber, a model explicitly specialized for autonomous vulnerability discovery and offensive cybersecurity operations, as NCT covered on June 23. The model’s stated purpose includes finding and exploiting software vulnerabilities without human intervention. It launched with no new export restrictions, no foreign-national access bars, and minimal regulatory pushback.

Anthropic’s Mythos 5 has comparable capabilities. It faced immediate export controls and blanket restrictions on foreign-national access.

The discrepancy, according to Axios, raises questions about whether the administration is enforcing safety policy or political preferences. The regulatory treatment of two models with similar risk profiles diverged sharply. One company’s model sailed through. The other’s triggered what amounts to a supply-chain crisis for every customer that depends on it.

For enterprise buyers, the implications are severe. Procurement decisions now carry regulatory risk that varies by vendor, not by capability. A company choosing between Anthropic and OpenAI is not just evaluating model performance. It is evaluating which company has better political relationships with the current administration. That calculus can change with every election cycle.

Eleven Days of Damage

The export controls directive landed on June 13. In the eleven days since:

Commerce Secretary Lutnick sent a formal letter to Anthropic threatening criminal penalties (June 16). Constitutional law scholars challenged the administration’s statutory authority to impose the restrictions (June 19). JPMorgan preemptively blocked Claude access for employees in Hong Kong (June 19). The administration reversed course and began framing Anthropic itself as a national security concern (June 20). CNN documented the absence of any coherent US AI regulatory framework (June 21). Chinese AI adoption accelerated as Zhipu’s GLM 5.2 and DeepSeek gained US corporate users who lost Anthropic access (June 22). Five Eyes intelligence agencies warned that autonomous AI threats are months away (June 22).

And on June 23: the first lawsuit, the NSA losing access, and documented evidence that the rules apply selectively.

The Agent Infrastructure Problem

Frontier AI models are no longer experimental tools that companies evaluate on quarterly refresh cycles. They are production infrastructure. Legion’s attorneys use Fable 5 to process legal documents. The NSA uses Mythos for cybersecurity operations. Agent platforms like OpenClaw route tasks through these models in real-time, with no manual fallback for most workflows.

When the government restricts access to a model, it does not just inconvenience users. It breaks production systems. The analogy is not banning a software product. It is cutting power to a data center. Everything downstream stops.

The selective enforcement compounds the problem. If regulatory treatment varies by company rather than by capability, then every agent deployment carries political risk. Enterprises cannot build reliable systems on infrastructure where access depends on which company has the right Washington relationships this quarter.

Legion’s lawsuit, the NSA’s access loss, and the Axios analysis each represent a different failure mode of the same policy. The lawsuit shows business operations breaking. The NSA story shows national security operations breaking. The selective enforcement shows the rule-of-law foundation breaking. All three happened on the same day.

The export controls are eleven days old. The first lawsuit, the first confirmed government-agency access loss, and the first documented evidence of selective enforcement all arrived in the same 24-hour window. The policy’s defenders argue it protects national security. The evidence so far suggests it is doing the opposite.