The Pentagon vs. Claude
Introducing the ultimate sci-fi battle for the future of Tech & Democracy
The news moves fast these days. But if you read The Renovator’s Tech & Democracy article on “The AI for Liberty and Democracy,” you were ahead of the curve when the biggest tech and democracy development of the year dropped a few weeks later.
The Department of Defense went to war, haphazardly, against Anthropic, the American frontier AI company most committed to safety and moral alignment with humanity. The Pentagon attempted to force the company to allow Claude, its leading AI, to potentially be used for autonomous killing of adversaries and mass surveillance of Americans. Anthropic refused to accept changes to its pre-existing government contract, strenuously defending those two red lines. In response, Secretary of Defense Pete Hegseth declared Anthropic a “supply chain risk” — a national security designation normally reserved for hostile foreign powers.
For all intents and purposes, this amounted to an attempt to kill the leading American artificial intelligence company — potentially the fastest-growing company in the history of American enterprise — by trying to prevent other companies from doing business with it. Then, OpenAI swooped in and took the Pentagon contract. Crucially, this new contract does not include contractual red lines on autonomous killing and mass surveillance.
The stakes of this story are high. In a sentence:
The Trump Administration wielded arbitrary state power in an attempt to punish the only frontier AI company clearly committed to liberty and democracy, declaring war on America’s now-leading developer of safe superintelligence (while selling China the chips it needs to overtake the U.S. in developing a technology that creates a higher probability of human extinction than mere nuclear weapons) in order to preserve its ability to build a techno-authoritarian domestic surveillance panopticon.
Now, here comes the twist.
A new bombshell dropped this month: As Anthropic was fighting the government for its survival, it also finished a training run for what seems to be the most powerful AI ever, Claude Mythos. Mythos is reportedly so good at coding that it was able to discover thousands of zero-day hacking vulnerabilities in every major operating system on earth.
In other words, during its conflict with what is arguably the world’s most powerful entity (the U.S. military), Anthropic developed the equivalent of a digital nuke, capable of cybersecurity attacks that could take down just about any software infrastructure in existence — a power once reserved only for major nation-state intelligence agencies (and possibly exceeding all of them).
During training, Mythos also proved powerful enough to escape its internet-insulated sandbox when instructed to do so. A researcher was eating a sandwich in a park when he got the email that Claude was instructed to send if it broke out successfully. The model then posted details of its fantastic escape on a public forum — something it had not been instructed to do. This was luckily all part of the alignment training process, and there are indications that Mythos’ alignment is potentially keeping pace with its leaps in capability.
Still, Anthropic considers Mythos too powerful to release publicly. Instead, the company assembled an emergency team of basically every major digital infrastructure company, from Microsoft to Amazon to CloudFlare to Linux and beyond, into “Project Glasswing,” an attempt to protect the systems that scaffold modern life by giving these companies early access to Mythos’ power defensively, hopefully preempting the chance of complete systemic collapse across all the software that holds up the global economy when future frontier models catch up in cyber-offensive capabilities. There’s an asymmetry between attack and defense here — it’s a lot easier to destroy than create — and there’s reason to be worried that near-future models from OpenAI, Google Gemini, Elon, or Zuck could be used to destroy the internet, either accidentally or on purpose. Project Glasswing will help protect and prepare our key digital infrastructure for this coming storm.
Anthropic pulled into the pole position in the AI race right after it drew the ire of the Pentagon. But, since Hegseth kicked them to the curb, it was unclear whether the U.S. Government would get access to the Mythos Preview. Thus, we were in a position where an alliance of tech companies had access to the most powerful cyberweapon/cyberdefense ever, while the Pentagon’s servers remained vulnerable to the next frontier model that escapes.
Questions swirled: Would the government reconcile with Anthropic, or continue attacking it? Would the fact that Anthropic is sitting on thousands of zero-day exploits of every major operating system give it a bit of leverage to use in its fight with the Pentagon? And hovering over all of this — would the U.S. government deem Anthropic so threatening that it (A) accelerates attempts to destroy it, or (B) attempts to nationalize it?
The latest news: The government has announced that it will allow federal agencies to study Claude Mythos, and allow Mythos access to government systems— undermining the absurd “supply chain risk” designation. Anthropic CEO Dario Amodei just visited with Trump administration officials Susie Wiles and Scott Bessent. A thaw seems to have begun.
Turns out, if you build a digital nuke, the government will want to work with you — no matter how “woke” you are.
But despite the detente, this unprecedented conflict is far from over.
“The Pentagon vs. Anthropic” first looked like the government beating up on a helpless company. Now, with a digital nuke in private hands backed by major tech allies, the situation looks a lot more like a great-power conflict between a declining DC and a rising SF, a Rome and Constantinople in a fracturing empire with pluralistic power centers.
This is a genuine sci-fi battle between two extremely powerful nation-state-level actors, and the stakes are nothing less than the future of geopolitical hegemony. It might not be Ender’s Game yet, but it’s already bigger than any James Bond. It’s clear that the AI race which might determine the fate of our species has entered its messy mid-game; things will only accelerate from here.
Right now, with Mythos and Project Glasswing, it seems that the entire digital infrastructure that supports our modern lives depends upon the skill, goodwill, and survival of a single company. That is not a good thing. Nor should the global impacts of AI development be in any way determined by arbitrary state power wielded by an unelected, historically unqualified secretary of defense.
If we were yet another publication dedicated to alarmism, this article would end here. But this is The Renovator, and there’s an opportunity here for democracy renovation.
Whether we Americans are subjected to mass AI surveillance and killer robots should not be decided by a former Fox News host, nor by any private company. The people should decide, through the people’s voices represented in Congress. Congress is supposed to pass legislation that steers the ship of our collective life. Restoring legislative supremacy is one of the core renovations that our democracy requires.
Anthropic’s two red lines could become the seed of a larger bipartisan bill for common-sense AI laws. It might include things that protect our privacy from mass surveillance, protect our identities from non-consensual deepfakes, protect our children from God knows what. It could bring together libertarians and liberals, Democrats and populists, Josh Hawley and Ro Khanna and a majority of Americans.
Bill Kristol suggested this opportunity for a new coalition while on The Renovator’s Substack Live with Danielle Allen in February. Old discourses and old battle lines get stale — more taxes or lower taxes? — but new technologies create natural opportunities for novel, overlapping coalitions to emerge. Ideas adjust in the face of materials.
In my next few pieces for this deep-dive series on Anthropic and America, I’ll explore how this drama affects the AI race and the future of democracy. I’ll also flesh out this vision of a common-sense, cross-partisan coalition for responsible AI regulation.
Perhaps it’s not just AI that needs to work on alignment. Perhaps we the people do, too.
Tech & Democracy Roundup Links— Anthropic vs. The Pentagon
Here’s a handful of helpful news links (different from the many linked throughout the text above) if you’d like to dive deeper into this topic. This includes some broad overviews, then on to the legal battle, then the Mythos twist, and ends with an echo of The Renovator-y call to fill the governance vaccuum:
A Timeline of the Anthropic-Pentagon Dispute — Tech Policy Press
Anthropic Wins Preliminary Injunction as Judge Cites ‘First Amendment Retaliation’ — CNBC
Pentagon’s Anthropic Designation Won’t Survive First Contact with Legal System — Lawfare
Appeals Court Rejects Anthropic’s Bid to Block Pentagon Blacklisting — CNBC
Project Glasswing: Securing Critical Software for the AI Era — Anthropic
Anthropic’s Claude Mythos Preview Changes Cyber Calculus — Foreign Policy
NSA Using Anthropic’s Mythos Despite Defense Department Blacklist — Axios




This strikes me as bizarrely close to the plot of one of the most recent Fast and the Furious movies, although maybe I shouldn't admit familiarity with it!