The AI That Found a 27-Year-Old Bug: Inside Anthropic’s Glasswing
I wanted to share this with you because something just shifted in the AI world, and it’s not another "better image generator" or a "faster chatbot."
Last week, Anthropic pulled back the curtain on something called Project Glasswing.
Most people are talking about the next version of Claude that can write better poetry or summarize longer PDFs. But while the public is focused on those creative tools, a small team has been building a specialized agent designed to do one thing: find the "invisible" cracks in the software that runs our entire civilization.
I’m talking about the code that manages your bank account, keeps the power grid online, and runs the operating systems we take for granted.
What I’ve noticed working with AI tools every day is that we’re moving away from "AI as an assistant" and toward "AI as a specialized operator." Glasswing is the clearest signal yet that this shift is here.
The "Secret" Model: Claude Mythos Preview
The core of Glasswing is a model you can’t use.
It’s called Claude Mythos Preview, and Anthropic is being very clear that they have no plans to release it to the general public.
Why? Because it’s too good at hacking.
In internal tests, Mythos Preview scored a 93.9% on SWE-bench Verified. To put that in perspective, the current top-tier models usually hover around 40-50% on these types of complex, agentic coding tasks. This isn't just a marginal improvement; it’s a generational leap in how an AI can navigate a codebase, understand terminal commands, and chain together complex actions to exploit a vulnerability.
But here’s the sincere part of the story: Anthropic isn't using it to break things. They're using it to fix them.
Finding the 27-Year-Old Needle in the Haystack
This is the part that really caught my attention. During the pilot phase, Mythos Preview found a bug in OpenBSD (a highly secure operating system) that had been sitting there, unnoticed, for 27 years.
Think about that. For nearly three decades, through millions of manual audits and automated tests, that flaw was just... there. Waiting.
Mythos Preview didn't just stumble upon it. It acted as an autonomous security researcher. It explored the code, tried different inputs, and eventually "thought" its way into finding the vulnerability.
If you’ve ever felt like the software you rely on is a bit of a "black box," you’re right. Most of the digital world is built on layers of legacy code that very few people truly understand anymore. Glasswing is like turning on a high-powered floodlight in a dark, dusty basement.
Why "Defensive AI" is the New Standard
You might find this useful if you've ever worried about the security of your own digital systems.
Traditionally, the "bad guys" have had the advantage in cybersecurity. They only have to find one hole to win, while the "good guys" have to plug every single one. Anthropic is trying to flip that script.
By partnering with companies like AWS, Google, and CrowdStrike, they are putting Mythos Preview to work scanning the most critical pieces of software on the planet. They are giving $100M in credits to open-source maintainers—the underfunded heroes who keep the internet's "plumbing" working—so they can use these tools to secure their code.
This is what I mean when I talk about Systems Over Tools.
Glasswing isn't just a tool; it's a defensive system. It represents a move toward a future where AI handles the "boring" but vital tasks of maintenance and security, freeing humans up to actually build new things.
Just Start Building (Safely)
I’ve been thinking about the problems we face as builders, and security is often the one we push to the back of our minds until something breaks.
We’re entering an era where you won't just "write code" anymore. You’ll be managing a fleet of specialized agents who write, test, and secure that code for you. Project Glasswing is just the first peek at what that looks like at a global scale.
The lesson for us is simple: don't just look for the "coolest" AI tool. Look for the one that makes your systems more durable.
If you’re looking for a way to put these types of agentic systems into practice in your own business without the usual trial and error, you might find the Builders Lab helpful. We provide the exact framework to help you create, track, and nurture your prospects, but we also dive deep into the technical systems that make it all work reliably. It’s designed to be a complete system you can lean on.
You can check it out here: https://www.thebuilderslab.pro/join
That's it for today. I hope this gave you a little more clarity on where the industry is headed.
Just start building.