The National Security Agency (NSA) tested Anthropic's Mythos AI model. In a red-team exercise on June 11, 2026, the AI — designed specifically for cybersecurity work — did what no human hacker has ever done. It broke into nearly everything.
Senator Mark Warner (D-VA), vice chairman of the Senate Intelligence Committee, confirmed the results during a closed-door briefing, relayed later through The Economist. Mythos, according to NSA Director General Joshua Rudd,
"broke into almost all of our classified systems, not in weeks, but in hours."
[Source: The Economist, "Donald Trump's blocking of Anthropic is capricious and chaotic," June 14, 2026]
Read that again. An AI, given access to the NSA's network for a defensive red-team test, penetrated nearly every classified system at the most powerful surveillance agency on earth — in a single afternoon.
That's the good news. It was a test. The bad news is what comes next.
Most Powerful AI In the World - In the Hands of The Untrustworthy
The same AI that cracked the NSA's defenses is now being deployed by the NSA itself — for offensive cyber operations against targets the government chooses. [Source: Financial Times, "Anthropic is helping the US National Security Agency deploy its Mythos AI model," June 4, 2026; TechCrunch, "NSA said to be readying Anthropic's Mythos for use in cyber operations," June 5, 2026]
And the same company that built Mythos — Anthropic — is the same company that weeks earlier refused to help the Heartland Institute design a website because the request might make climate-skeptic messages "more polished and persuasive." [Source: Document #1619, The Blaze, "Claude shows the future: No AI for dissenters," June 18, 2026]
The connection is not accidental. It is the logical endpoint of a pipeline I traced in AI Refused! - They Fired a Warning Shot: first they control what the AI says, then they control who gets to use it, and finally — inevitably — they weaponize it against those they've already excluded.
The Whiplash Timeline
To understand how we got here, follow the dates. They tell a story the press releases won't.
February 2026: The Pentagon moves to designate Anthropic a "supply chain risk" — the same label applied to foreign adversaries like Huawei and ZTE. The dispute began when Anthropic refused to allow its AI to be used for mass domestic surveillance and fully autonomous weapons systems. [Source: TechSpot, "Anthropic is blacklisted by the Pentagon and being used by the NSA at the same time," June 6, 2026; TechCrunch, "It's official: The Pentagon has labeled Anthropic a supply chain risk," March 5, 2026]
March 2026: The Pentagon makes it official. Defense contractors are barred from using Anthropic's technology. Anthropic sues. A federal judge temporarily blocks the designation, but an appeals court allows it to stand. The company that built the most capable AI model on earth is now formally considered a national security threat by the US military. [Source: TechSpot, June 6, 2026]
April 2026: While the Pentagon fights Anthropic in court, the NSA — which sits *inside* the Department of Defense — secretly begins using Mythos. Axios breaks the story. "The National Security Agency is using Anthropic's most powerful model yet, Mythos Preview, despite top officials at the Department of Defense insisting the company is a 'supply chain risk,'" two sources tell Axios. [Source: Security Affairs, "The US NSA is using Anthropic's Claude Mythos despite supply chain risk," April 21, 2026; Axios, "NSA using Anthropic Mythos despite Pentagon concerns," April 19, 2026]
April 17, 2026: Anthropic CEO Dario Amodei meets White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent. The White House calls it "productive and constructive." The relationship is thawing. [Source: TechCrunch, "Anthropic's relationship with the Trump administration seems to be thawing," April 18, 2026; Axios, April 17, 2026]
June 4-5, 2026: The Financial Times reports that Anthropic has deployed half a dozen engineers *inside the NSA* to help the agency use Mythos for offensive cyber operations. One source close to the arrangement tells the FT the model would be useful for infiltrating networks in countries like China and Iran. Reuters reports that tensions between Anthropic and the White House are easing ahead of the company's planned IPO. [Source: TechCrunch, "NSA said to be readying Anthropic's Mythos for use in cyber operations," June 5, 2026; Reuters, "Blacklisted AI company Anthropic, White House ease tensions ahead of IPO," June 5, 2026]
June 9, 2026: Anthropic publicly launches Fable 5 and Mythos 5 — the most powerful AI models ever released to consumers. The company had previously described Mythos as too powerful in cybersecurity to release publicly. [Source: Forbes, "Anthropic Disabled Fable 5 And Mythos 5 After A U.S. Export-Control Order," June 16, 2026]
June 11, 2026: The NSA red-team test. Mythos penetrates "nearly all" classified systems in hours. [Source: The Economist, June 14, 2026]
June 12, 2026: The Commerce Department issues an emergency export control directive, ordering Anthropic to suspend access to Fable 5 and Mythos 5 by any foreign national, anywhere in the world — including foreign national Anthropic employees. Anthropic disables both models globally. [Source: Ars Technica, "Anthropic shuts down Fable, Mythos models following Trump admin directive," June 13, 2026; Forbes, June 16, 2026]
June 15, 2026: Anthropic staffers rush to Washington to meet with administration officials. The company disputes the severity of the alleged jailbreak that triggered the shutdown. White House AI adviser David Sacks claims Anthropic refused to fix the issue; Anthropic says it was never given details. [Source: CNBC, "Anthropic to meet with Trump administration over Mythos dispute," June 15, 2026; Forbes, June 16, 2026]
June 21, 2026: Reports indicate the Trump administration is moving toward renewing its relationship with Anthropic. The company that was labeled a "supply chain risk" four months ago is now preparing to supply the US government's most powerful cyberweapons.
The whiplash is the point. The government can brand you an enemy in the morning and a partner by evening — if you have something they want.
The Connection You're Not Supposed to Make
Here is the thread that connects every piece of this story.
In June 2026 — the same month Mythos was cracking NSA systems and Anthropic engineers were embedding themselves inside America's signals intelligence apparatus — Anthropic's Claude refused to help the Heartland Institute design a website. [Source: Document #1619]
The request was legal. It was not violent, fraudulent, or harassing. It asked for *user interface suggestions*. Anthropic's AI refused because making the Heartland Institute's website "more polished and persuasive" would have a negative "downstream effect" on climate discourse. [Source: Document #1619]
The AI company that decides who is "safe" to use its productivity tools — that judges the ideological acceptability of every request — is now building the US government's cyber arsenal.
Think about what that means.
Anthropic has demonstrated it will deny access to its AI based on the political views of the user. The Heartland Institute is a conservative think tank. The refusal was ideological — an exercise in gatekeeping dressed as safety. [Source: Document #1619]
The same company is now placing its engineers inside the National Security Agency. The same company is building the AI model that can break into "nearly all" classified systems in hours. [Source: The Economist, June 14, 2026]
The same safety protocols that blocked a lawful website design request are not going to stop the offensive cyber operations that Mythos is now being trained to conduct.
The guardrails are for you. They are not for the government.
As I wrote in AI Refused! - They Fired a Warning Shot: "First they control what the AI says, then they control who gets to use it." The third step is now in motion. They weaponize it.
The Danger: Who Gets Targeted?
The NSA's clandestine signals intelligence apparatus has a long and well-documented history of overreach. Section 702 of the Foreign Intelligence Surveillance Act was originally sold as a tool for targeting foreign terrorists. In practice, it has been used to collect the communications of American citizens without warrants — including Jan6'er David Moerschel, members of Congress, journalists, and political activists.
During the "Arctic Frost" investigation, Section 702 queries were used by Special Counsel Jack Smith. When Representative Andy Biggs pressed Attorney General Pam Bondi on the legal predicate for using a foreign intelligence authority in that investigation, Bondi could not answer. The matter was referred to her office as an active investigation. Queries had swept up members of Congress, congressional staff, journalists, and other US persons not suspected of acting as foreign agents. [Source: Document #1411, "Rep. Biggs Presses DOJ on FISA 702 and Arctic Frost Investigation," February 11, 2026]
That is the existing surveillance apparatus. Now imagine it supercharged with an AI that can crack "nearly all" classified systems in hours.
The NSA is already using Mythos for offensive cyber operations. Anthropic engineers are embedded inside the agency to help guide its use. [Source: TechCrunch, June 5, 2026; Financial Times, June 4, 2026]
If Mythos can break into an NSA-classified system in hours, what can it do to the private communications of a political dissident? What can it do to the encrypted messaging of a journalist investigating government overreach? What can it do to the networks of a think tank the government has decided is a threat?
The AI doesn't care about warrants. It doesn't care about legal predicate. It finds vulnerabilities and exploits them. That's what it was designed to do.
And the company that decides who gets access to that power is the same company that refused to help a conservative think tank design a website because their message was ideologically disfavored. The pattern is formed already. We are the targets.
The Ideological Filter Meets the Surveillance State
The convergence is almost complete.
The AI companies have built systems that judge the ideological acceptability of every request — refusing service to organizations whose views fall outside the approved spectrum. [Source: Document #1619]
The same companies are now embedding their technology inside the most powerful surveillance state in human history. The NSA — an agency that has conducted warrantless surveillance on American citizens, that has been caught spying on Congress, that has its own Inspector General say it cannot accurately count how many times it has violated its own rules — is getting the most powerful cyberweapon ever built. [Source: Document #1411]
The Pentagon labeled Anthropic a "supply chain risk" — the same label applied to Huawei — because Anthropic initially refused to let the military use its AI for mass domestic surveillance and autonomous weapons. [Source: TechSpot, June 6, 2026; TechCrunch, March 5, 2026]
But the NSA is using it anyway. The White House is negotiating with Anthropic's CEO. The export controls that briefly shut down Mythos and Fable were not about protecting civil liberties — they were about keeping the technology out of foreign hands. A licensing regime for frontier AI. [Source: Fortune, "Make no mistake: the U.S. now has a licensing regime for frontier AI," June 16, 2026]
The government doesn't want to stop Mythos. It wants to own it.
And when the government owns the AI — and the AI already decides who is "safe" to use its tools — the circle closes. The ideological filter becomes the surveillance state's gate. The same system that denies you access to AI because of your political views is the same system that can crack your encrypted communications because the government flagged you as a target.
This is not paranoia. This is the trajectory of every documented event.
The Call: Technological Independence Is Survival
In "AI Refused! - They Fired a Warning Shot," I wrote that the alternative to this system exists. Local AI models — running on personal hardware, not corporate servers — can't be switched off by a trust-and-safety team. They don't require digital ID to operate. They don't have ideological filters baked in by San Francisco executives.
That truth is even more urgent today.
The AI that refused the Heartland Institute's website request is the same architecture — the same company, the same engineering team, the same corporate leadership — that is now building the NSA's offensive cyber capabilities. The ideological gatekeeping and the surveillance weapon are products of the same system.
If you rely on corporate AI for anything sensitive — anything that might be disfavored by the political establishment, anything that might attract the attention of a surveillance state armed with Mythos-class capabilities — you are trusting the same people who decide whether your website design request is ideologically permissible.
You should not make that bet.
The lifeboats are local AI, encrypted communication, and decentralized technology. They exist. But the window to board them is closing.
The NSA tested Mythos. It cracked nearly everything in hours.
The question is not whether the government will use that capability. It already is. The question is whether you will have any defense when it's turned on you. Exactly how many hours after Donald Trump leaves office, do you think, before the bad guys flip the switch and turn it on us?
---
Sources
Document #1619: "Claude shows the future: No AI for dissenters" — The Blaze, Donald Kendal, June 18, 2026
Document #1411: "Rep. Biggs Presses DOJ on FISA 702 and 'Arctic Frost' Investigation" — Congressional Hearing Transcript, February 11, 2026
The Economist: "Donald Trump's blocking of Anthropic is capricious and chaotic" — June 14, 2026 (includes NSA Director's statement that Mythos "broke into almost all of our classified systems, not in weeks, but in hours")
Ars Technica: "Anthropic shuts down Fable, Mythos models following Trump admin directive" — June 13, 2026
Forbes (Anisha Sircar): "Anthropic Disabled Fable 5 And Mythos 5 After A U.S. Export-Control Order. Here's What Happened" — June 16, 2026
CNBC: "Anthropic to meet with Trump administration over Mythos dispute" — June 15, 2026
TechCrunch: "Anthropic's relationship with the Trump administration seems to be thawing" — April 18, 2026
TechCrunch: "NSA spies are reportedly using Anthropic's Mythos, despite Pentagon feud" — April 20, 2026
TechCrunch: "NSA said to be readying Anthropic's Mythos for use in cyber operations" — June 5, 2026
Financial Times: "Anthropic is helping the US National Security Agency deploy its Mythos AI model" — June 4, 2026 (reported via TechCrunch, TechSpot)
Security Affairs: "The US NSA is using Anthropic's Claude Mythos despite supply chain risk" — April 21, 2026
TechSpot: "Anthropic is blacklisted by the Pentagon and being used by the NSA at the same time" — June 6, 2026
Reuters: "Blacklisted AI company Anthropic, White House ease tensions ahead of IPO, sources say" — June 5, 2026
Axios: "NSA using Anthropic Mythos despite Pentagon concerns" — April 19, 2026 (reported via Security Affairs, TechCrunch)
Fortune: "Make no mistake: the U.S. now has a licensing regime for frontier AI" — June 16, 2026
---
Simon Ambrose is an investigative journalist covering the intersection of technology, surveillance, and civil liberties. This article is the second in a series examining the converging systems of AI censorship and state power.