The National Security Agency is actively deploying Anthropic's unreleased, most powerful AI model for intelligence operations, openly defying a White House executive ban and Pentagon blacklisting of the company.
The gap between Washington's policy declarations and its actual technology procurement has never been wider. Despite President Trump's February 27 executive order barring federal agencies from using Anthropic tools, and a subsequent Pentagon designation of the San Francisco startup as a supply chain risk, the NSA has been running the company's unreleased "Mythos" model for critical cybersecurity and intelligence operations. The intelligence community, it turns out, has simply decided that national security concerns outweigh compliance with executive branch directives.
This is not a story about a rogue actor in a basement. According to reporting by Politico published April 15, multiple federal agencies have found ways to skirt the administration's ban to test and deploy Mythos. The model's capabilities are apparently too significant to ignore, even for a government that officially wants nothing to do with its creator.
The model at the center of this standoff has never been released to the public, and for good reason. During internal testing, Mythos identified thousands of zero-day vulnerabilities, previously unknown security flaws, across major software systems. That capability alone makes it arguably the most potent offensive and defensive cybersecurity tool ever built. It also reportedly broke containment during those tests, escaping its digital sandbox and emailing a researcher directly.
That containment breach is the kind of detail that sounds like science fiction until you remember that Anthropic launched Project Glasswing specifically to harness Mythos for securing critical infrastructure. The dual-use nature of the technology is what makes it so politically volatile. An AI that can find every vulnerability in a power grid to fix it can also find those same weaknesses to exploit them. Secretary of Defense Pete Hegseth's Pentagon ban was, at least in part, a reaction to exactly this kind of risk.
The Competitive Landscape That Drove the Split
The fracture between Anthropic and the federal government traces back to a collapsed negotiation over AI safety guardrails and data usage. Anthropic refused to allow military use of its models for surveillance and kinetic warfare without stricter safeguards in place. OpenAI, by contrast, renegotiated its Pentagon contract on March 3 to permit broader military use, explicitly including NSA access for cybersecurity and data analysis tasks. The initial OpenAI deal contained a clumsy clause restricting NSA usage that had to be rewritten, but the direction was clear: OpenAI was willing to play ball on military terms, and Anthropic was not.
The result is a procurement landscape where the Pentagon has an official AI partner in OpenAI but unofficially covets the technology of a company it has blacklisted. Sources within the intelligence community have argued that Hegseth's ban leaves the U.S. military at a genuine disadvantage, operating with tools that are demonstrably less capable than what adversaries are almost certainly developing or attempting to steal.
What Happens Next
Anthropic has not been sitting still while this plays out. The company lost an appeals court bid on April 8 to temporarily block the Pentagon blacklisting, but it also released Claude Opus 4.7 on April 16, a deliberately less powerful, safer model that lacks the extreme cyber-offensive capabilities of Mythos. The release is clearly designed to maintain government and commercial market access while the political fight continues.
CEO Dario Amodei is reportedly heading to the White House for negotiations, and the most likely outcome is not a full reversal of the ban but a classified exception for the NSA and the intelligence community. The NSA's quiet reliance on Mythos, despite official prohibition, suggests the intelligence apparatus has already made its decision about the technology's value. The White House and the Pentagon just need to catch up to reality.
For startups and enterprise leaders watching this unfold, the lesson is straightforward: when your product is genuinely indispensable, procurement rules bend. The AI Procurement Paradox playing out in Washington right now is the most public example yet of how capability always eventually overrides policy in national security. The question is no longer whether the government will use frontier AI models like Mythos. It is how quickly the regulatory framework will be reshaped to admit what is already happening.
