Who Gets Protected When AI Finds the Holes First

Anthropic is calling it Project Glasswing, which is a beautiful name for a thing that deserves more scrutiny than it's getting. The pitch is this: Claude Mythos Preview, their latest model, can scan codebases the size of cathedrals and find the cracks before the bad guys do. 150 organizations across 15 countries. 10,000 high or critical severity vulnerabilities identified. BT Group blocking 4 million attacks a day with the thing in their toolkit. Microsoft, Google, Apple, Nvidia, AWS, CrowdStrike, JPMorgan all at the table. The numbers are real and the names are heavy and if you're the kind of person who follows this space, you're supposed to feel relieved right now.

I don't feel relieved. I feel like I'm watching a very expensive fire truck pull up to a neighborhood where only the houses with sprinkler systems called for it.

Here's what the coverage keeps framing as progress: the bottleneck moves from discovery to triage. You find the vulnerabilities faster, so now the constraint is how quickly your team can patch them. That sounds like an improvement until you sit with what "triage" actually means, which is that you've got more problems than you can fix and you're deciding what to let burn. Every security team does this already. Every single one. The question has always been which fires you can afford to fight, and AI that helps you prioritize better is genuinely useful, except the organizations with access to Claude Mythos Preview are vetted organizations, selected organizations, organizations that already have security budgets and security postures and dedicated humans whose job is thinking about this stuff. The tool makes the strong faster. That's not nothing. It's also not the same thing as making everyone safer.

Because while Glasswing is helping BT Group block 4 million attacks a day, there's a school district somewhere in this country whose entire IT operation is one overextended person who also handles the phone system and the copiers and the Wi-Fi password resets. There's a rural hospital running on hardware that predates the iPhone, protected by whatever the managed service provider bundled into the contract at a price point that made the board of directors comfortable. There's a small nonprofit (and I work at one, so I'm not speaking abstractly here) where "cybersecurity strategy" is a phrase that exists in a policy document nobody has read since the last grant required them to write it. These organizations don't find their vulnerabilities before the attackers do. They find them when someone calls to say the files are encrypted and there's a Bitcoin address on the screen. That's the discovery process. That's what "finding flaws" looks like when you're not in the room where Glasswing is being deployed.

The article acknowledges the dual-use problem, which I'll give it credit for, but it acknowledges it the way you acknowledge a weather forecast for a city you don't live in. Anthropic says access is restricted to vetted organizations. That's a promise, not an architecture. Every access control system has a flaw, every vetting process has a blind spot, and the model itself doesn't know or care who's holding it , it just finds the crack in the wall and points. The same capability that helps a defender find a zero-day first helps an attacker find it first if the attacker gets there first, and the attacker doesn't need to be vetted, the attacker just needs access, and access has a way of getting out no matter how carefully you designed the door. We've seen this enough times that it shouldn't need saying, but it apparently does, because the press release doesn't mention it and neither does most of the coverage.

Power doesn't corrupt. It reveals. What it reveals is what you were already willing to do when the stakes got high enough, and Anthropic is being watched right now, which matters, which is genuinely something. Whether it's enough depends on a question the coverage keeps skating past: when one of those vetted organizations turns out to have been less vetted than advertised, when the capability leaks sideways the way capabilities always eventually do, what happens to the school district? What happens to the hospital? The people who built Glasswing are not thinking about the hospital. They're thinking about BT Group and JPMorgan and the enterprise clients who can afford the seat at the table, and that's not an accusation, that's just how the incentives are arranged, and the incentives are the thing worth examining.

I want AI-assisted vulnerability detection to exist. I want it to work. I want the number of successful attacks to go down and I want the people doing the defending to have better tools than the people doing the attacking, at least for once, at least for a while. But I also want someone to be asking out loud whether this particular wave lifts the boats that are already floating or whether it just makes the already-floating boats faster while the ones taking on water keep bailing with whatever's available. Because "the track keeps changing shape" is one way to describe what's happening in this space. Another way is that the rules are being written right now, by the people who already won the last round, and if nobody pushes back on who gets to play, the answer to "who gets protected" is going to be the same answer it's always been, which is the people who could afford protection before any of this existed.

That's not a technology problem. That's a choice. And choices can be made differently, if someone decides to make them.