Here’s the question nobody’s asking: If an AI model can find thousands of zero-day vulnerabilities in weeks, what does that say about the ones still hiding in your supply chain right now?
Anthhropic unveiled Mythos this week—a frontier model so capable that the company initially kept it under wraps with code names like “Capybara” until someone left the blueprints in an unsecured data lake. (More on that later.) The model is now in the hands of exactly 52 organizations: 12 strategic partners running Project Glasswing, plus 40 others getting preview access. And the headlines have been predictable: “AI Finds Thousands of Zero-Days!” “Revolutionary Security Tool!” The usual.
But here’s what matters: Mythos is caught in a collision between Silicon Valley’s security ambitions and a government that views AI labs as potential national security threats.
The Numbers Look Staggering—Until You Read the Fine Print
Anthhropic claims Mythos identified “thousands of zero-day vulnerabilities, many of them critical,” with some dormant for one to two decades. That’s genuinely impressive. Partners include Amazon, Microsoft, Apple, Cisco, and Palo Alto Networks—not exactly fringe players.
“Mythos identified ‘thousands of zero-day vulnerabilities, many of them critical,’ with some dormant for one to two decades.”
But—and this is a significant but—we don’t know what “thousands” really means. Is it 2,000? 10,000? We don’t know how many were actually exploitable versus theoretical. We don’t know the false-positive rate. Anthropic’s leak materials mentioned the model “could potentially pose a cybersecurity threat if weaponized by bad actors to find bugs and exploit them.” That’s a corporate way of saying: this thing is powerful enough to be dangerous.
The company also didn’t train Mythos specifically for cybersecurity. It’s a general-purpose model that happens to be very good at reasoning about code. That’s actually telling—it suggests Mythos’s power comes from raw intelligence rather than specialized tuning.
Why the Pentagon Memo Changes Everything
Here’s where the story gets sharp. Anthropic is currently in a legal battle with the Trump administration after the Pentagon labeled it a “supply-chain risk.” The Pentagon’s complaint centers on Anthropic’s refusal to allow autonomous targeting or surveillance of U.S. citizens.
So while Anthropic’s team is in “ongoing discussions” with federal officials about Mythos, those discussions are happening in the shadow of active litigation. It’s hard to imagine a more fraught dynamic: the company wants government blessing to deploy a powerful AI model into critical infrastructure, but the government has already flagged the company as a risk. Anthropic claims the Pentagon decision was political retaliation. The Pentagon sees it as risk management.
Who’s right? That’s a different argument. But the timing is toxic.
Is There Actually a Security Benefit Here?
Yes. Scanning legacy and open-source code for vulnerabilities is genuinely useful work. One-to-two-decade-old bugs in widely-used libraries? Those are real attack surfaces that conventional tools have missed. If Mythos can systematically find those, defenders win.
The partner commitment to share findings across the industry is solid too. That’s how you turn a capability into a public good—or at least, that’s the stated goal.
What’s less clear: whether this model’s power is proportional to its risk. Anthropic’s own leaked materials acknowledged that the same reasoning skills that find vulnerabilities could find them to exploit, not fix. That’s not a knock against the company’s intentions. It’s just basic information security: powerful tools have dual-use potential.
The Data Breach Nobody Wants to Talk About
Let’s pause on the positive press for a second. Last month, Anthropic leaked nearly 2,000 source code files and over 500,000 lines of code via a mistake in launching Claude Code version 2.1.88. Then, while trying to clean up, it accidentally triggered thousands of GitHub takedowns.
A leaked Mythos roadmap document. A code-upload disaster. These aren’t unrelated. They’re evidence that Anthropic’s operational security—the actual practice of keeping sensitive things secret—doesn’t match its rhetoric about being a trustworthy steward of powerful AI.
That matters when you’re asking the government to let you deploy a model that could theoretically be weaponized into critical infrastructure.
The Real Test Isn’t the Technology
Mythos will probably work as advertised. The vulnerabilities it finds will probably save real systems from real attacks. That’s the easy part of the story.
The hard part is whether an AI lab that’s mid-lawsuit with the Pentagon, that’s already had multiple operational security failures, and that’s building increasingly powerful models, can convince policymakers that this is worth the risk. Right now, it’s a credibility problem masquerading as a technology problem.
The 52 organizations with access will learn a lot. The industry might benefit. But Project Glasswing’s real value won’t be measured in zero-days found. It’ll be measured in whether it changes how Washington thinks about Anthropic. And given the current legal backdrop, that’s a much steeper climb than scanning code repositories.
🧬 Related Insights
- Read more: Anthropic’s Back-to-Back Leaks Hand Rivals Claude Code’s Blueprint
- Read more: EU AI Act Hands SMEs a Real Edge Over Big Tech
Frequently Asked Questions
What does Anthropic’s Mythos model actually do?
Mythos is a general-purpose AI model trained to scan codebases—both proprietary and open-source—for software vulnerabilities. Anthropic claims it’s identified thousands of zero-day bugs since its limited launch. It wasn’t specifically trained for security work; it’s an application of the model’s broader reasoning skills.
Will Mythos be available to everyone?
No. Only 52 organizations have access: 12 strategic partners in Project Glasswing (including Amazon, Microsoft, Apple, Cisco) plus 40 additional organizations in a broader preview. Anthropic hasn’t committed to public release.
Why is the Pentagon concerned about Anthropic?
The Trump administration’s Pentagon labeled Anthropic a supply-chain risk because the company refused to allow autonomous targeting or surveillance of U.S. citizens. The two are currently in active litigation. Anthropic’s ongoing talks with federal officials about Mythos deployment happen against this hostile backdrop.
