💻 technology

Spy Agency Sneaks Around Trump's AI Ban

The NSA is quietly using Anthropic's new Mythos AI model for cybersecurity work, despite Trump's February order banning all federal agencies from touching the company's tech. While the White House and Pentagon are still locked in legal warfare with Anthropic, America's top signals intelligence agency apparently didn't get the memo - or chose to ignore it.

The NSA (National Security Agency) is actively deploying Anthropic's Mythos Preview model, a specialized AI designed for computer security tasks, according to two sources familiar with the arrangement. This puts the intelligence agency in direct conflict with a Trump administration executive order from February 2026 that explicitly banned federal agencies from using Anthropic's services. The company released Mythos Preview in early April 2026, marketing it as exceptionally capable at cybersecurity operations - precisely the kind of work the NSA does every day. The context here is a months-long brawl between Anthropic and the Pentagon that started when the AI company refused to remove certain safety guardrails during military contract negotiations. Anthropic's position was simple: we built these models with specific ethical boundaries, and we're not stripping them out just because the Department of Defense wants unrestricted access. Trump's response was to designate Anthropic a supply chain risk and order a government-wide boycott. The company fought back with lawsuits in two separate courts in March 2026. The legal battle has produced mixed results. One federal court granted Anthropic a preliminary injunction temporarily blocking the supply chain risk designation. But another court denied the company's motion to lift the label entirely, leaving Anthropic in regulatory limbo. The Pentagon filed its own response defending the designation, setting up what could be years of litigation. Meanwhile, Anthropic CEO Dario Amodei showed up at the White House on April 17, 2026, for what officials later called a productive and constructive meeting with Chief of Staff Susie Wiles and other administration figures. The topic? Reportedly Mythos itself. When reporters asked Trump about the meeting, he claimed to have no idea it happened - a classic Trump move when something politically awkward is brewing. The NSA is apparently one of roughly 40 organizations Anthropic granted early access to Mythos Preview, and sources say the model is seeing expanding use within the Department of Defense more broadly. What makes this particularly messy is that the NSA, technically part of the Department of Defense, appears to be operating under a different playbook than the rest of the federal government. Either the intelligence community carved out an exception for itself, or the NSA decided that national security imperatives trump executive orders. Given the agency's track record of doing whatever it believes necessary for signals intelligence, the latter seems more likely. Anthropic's cybersecurity-focused AI is exactly the kind of tool the NSA would consider mission-critical, regardless of what political drama is unfolding in Washington.

My Take

The NSA's decision to use Mythos anyway reveals something important about how real national security works versus how political theater works. Trump can issue all the executive orders he wants, but when an intelligence agency believes a tool is critical for protecting American interests, bureaucratic niceties go out the window. The NSA isn't exactly known for asking permission - this is the agency that built a global surveillance apparatus before most Americans even knew what metadata meant. What's actually happening here is a proxy war over AI governance. Anthropic is fighting for the principle that AI companies should retain some control over how their models get deployed, especially for military and intelligence purposes. The Pentagon wants unfettered access - no safety rails, no ethical guardrails, just raw capability. The NSA's end-run around Trump's ban suggests the intelligence community agrees with Anthropic that specialized, purpose-built models with appropriate constraints are more valuable than general-purpose systems with all the safeties removed. The timing of Amodei's White House visit is telling. You don't schedule a meeting with the Chief of Staff while simultaneously suing the federal government unless someone is trying to cut a deal. My guess? Anthropic is offering the administration a face-saving compromise: restricted access to Mythos for specific national security uses, with the company maintaining some oversight, in exchange for dropping the supply chain risk designation. Whether Trump's ego will allow him to back down from his February ultimatum is another question entirely.

What Happens Next

The NSA's defiance creates an impossible position for the White House. If Trump enforces his own executive order, he risks crippling the intelligence community's cybersecurity capabilities at a moment when state-sponsored hackers from China, Russia, and Iran are increasingly using AI-powered tools. If he backs down, he looks weak and undermines his own Department of Defense's legal position in the Anthropic lawsuits. The smart money is on a quiet carve-out: the administration will claim that intelligence activities fall under a different legal framework than general federal procurement, allowing the NSA to continue using Mythos while maintaining the official ban for civilian agencies. Watch for Anthropic to leverage this split to negotiate better terms. If the NSA is publicly using Mythos and finding it indispensable, other Defense Department units will demand access too. The Pentagon's hardline position becomes untenable when its own intelligence arm is contradicting it in practice. Expect a settlement within 90 days that lets the Defense Department save face while giving Anthropic most of what it wanted: AI models deployed for national security with company-maintained ethical boundaries intact. The wildcard is Congress. If lawmakers discover that the NSA ignored a presidential directive, expect hearings. Republicans will be torn between supporting Trump and supporting the intelligence community. Democrats will weaponize this as evidence of chaos in the administration. Meanwhile, China and Russia are watching closely to see if American political dysfunction will slow US military AI development enough to give them a competitive window. They're not waiting for Washington to sort out its internal drama.

What History Tells Us

This standoff echoes the encryption wars of the 1990s and 2010s, when tech companies clashed with intelligence agencies over backdoor access to secure communications. In 2016, Apple refused the FBI's demand to unlock the San Bernardino shooter's iPhone, arguing that creating a backdoor would compromise security for all users. Apple won that battle in the court of public opinion, even as the FBI found other ways to access the device. The parallel here is striking: Anthropic is essentially arguing that removing AI safety guardrails to satisfy Pentagon demands would create vulnerabilities that outweigh the short-term intelligence value. There's also historical precedent for intelligence agencies simply ignoring executive branch directives they consider operationally harmful. During the Church Committee investigations of 1975, Congress discovered that the NSA and CIA had conducted domestic surveillance programs for decades despite presidential orders theoretically limiting such activities. The agencies justified their actions as necessary for national security, and most of the programs continued in modified form. If history is any guide, the NSA will keep using whatever tools it deems essential, and the political class will eventually ratify the fait accompli.