Anthropic just handed the keys to its most dangerous AI model to NATO, the European Union’s cybersecurity agency, Samsung, SK Hynix, and the companies that keep the lights on across more than 15 countries. The model is called Claude Mythos. It is the same system Anthropic has described as so capable it can find and exploit thousands of zero-day vulnerabilities in a matter of weeks — the kind of tool that, in the wrong hands, is indistinguishable from a cyberweapon. On June 2, the company decided the right hands include power grids, water utilities, hospitals, and a military alliance.
This is the part of the AI story nobody at the launch events wants to say plainly: the most powerful frontier models are no longer just writing your emails. They are quietly becoming critical infrastructure themselves — and the only thing separating “defending the grid” from “owning the grid” is a set of safeguards Anthropic is writing on the fly.
What Anthropic Actually Announced
The expansion runs through Project Glasswing, Anthropic’s joint industry initiative to find and fix critical software flaws using AI. According to TechCrunch and Cybersecurity Dive, Anthropic is widening access to roughly 150 new organizations across more than 15 countries. The headline names tell you exactly how high the stakes are: identity-security firm Okta, South Korea’s Samsung, SK Hynix, and SK Telecom, NATO, and ENISA — the EU’s cybersecurity agency.
The target is explicit. Anthropic says Glasswing is aimed at critical infrastructure in power, water, healthcare, and communications — sectors where, in the company’s own framing, a single successful cyberattack could affect 100 million people. And the engine doing the work is Claude Mythos, which Anthropic has called its most powerful model yet: a general-purpose frontier system that, per Anthropic’s own red-team disclosures, can “surpass all but the most skilled humans at finding and exploiting software vulnerabilities.” Since the project launched, Anthropic says partners have already surfaced more than 10,000 high- or critical-severity security flaws.
Say the Quiet Part: This Is a Dual-Use Weapon With a Friendly Logo
Here is the uncomfortable truth buried under the press-release language about “securing critical software for the AI era.” A model that can find 10,000 vulnerabilities to patch them is, by definition, a model that can find 10,000 vulnerabilities to exploit them. The capability is identical. Only the intent changes. Anthropic knows this — it is why the company itself flagged an earlier version of this exact capability as too dangerous to release openly.
So the real news isn’t that Anthropic built a vulnerability scanner. It’s that Anthropic has decided the safest place for a cyber-offensive-grade model is inside a controlled coalition of trusted operators — and that it gets to define who is trusted. NATO is in. ENISA is in. Samsung and SK Hynix, two of the most strategically important chipmakers on Earth, are in. Anthropic isn’t just selling a product. It’s assembling a security bloc, and writing the membership rules.
The Second-Order Effect: Anthropic Becomes Infrastructure
Follow the consequences, not the announcement. If the power companies, water utilities, and telecom backbones of 15-plus countries start relying on Mythos to find the holes in their software before attackers do, then Anthropic stops being a vendor and becomes a dependency. You cannot easily rip out the system that knows where all your unpatched flaws live. That is the stickiest kind of lock-in there is — and it arrives wrapped in the most defensible justification imaginable: national security.
The timing is not subtle either. This expansion lands days after Anthropic confidentially filed for an IPO at a reported valuation near $965 billion, on a revenue run-rate that has it overtaking OpenAI. A company about to ask public markets to believe it is the most valuable startup ever to list now has a second story to tell investors: we are not just an AI lab, we are the security layer for the Western world’s critical infrastructure. That narrative is worth more than any benchmark score.
Who Gets Hurt — And What Happens When the Rivals Catch Up
Anthropic has been refreshingly blunt about the scary part: it expects other AI companies to develop models as capable as Mythos Preview soon, which is precisely why it is racing to establish safeguards now. Read that again. The company is essentially admitting that this class of capability is about to proliferate, and that its “controlled access” model is a stopgap, not a moat. When OpenAI, Google, or a Chinese lab ships a Mythos-equivalent with looser controls — or when one leaks, as Anthropic’s own models reportedly have before — the genie does not go back in the bottle.
The losers in the short term are the organizations not invited into Glasswing. Critical infrastructure operators outside the trusted circle now face an asymmetry: their adversaries may soon have AI that can autonomously hunt zero-days, while their defenders are still doing it the manual way. “Voluntary coordination with industry,” the phrase governments and Anthropic both keep using, is a polite way of saying there is no enforcement, no floor, and no guarantee the same capability stays defensive once it spreads.
The Verdict
Anthropic is doing something genuinely useful and genuinely unsettling at the same time, and pretending otherwise is the mistake. Finding 10,000 critical flaws in the world’s most important software before attackers do is real, defensible work. But the company has also confirmed what the cyber community has whispered for two years: frontier AI is now good enough to attack critical infrastructure, and the line between Anthropic-the-defender and Anthropic-the-single-point-of-failure is drawn entirely by Anthropic. Project Glasswing isn’t a cybersecurity announcement. It’s the moment an AI lab quietly became part of the grid — and asked us to trust that “voluntary safeguards” will hold while every competitor races to build the same weapon. They probably won’t. The honest takeaway is that we are now depending on the good intentions of private AI companies to keep the power on. That should make you grateful and nervous in exactly equal measure.