
We are here, my babies. A canon event.
Anthropic’s latest model, Claude Mythos Preview, has been kept off the public market. Instead of a splashy launch, we got Project Glasswing, which gives 12 launch partners and 40 additional organizations access to Mythos Preview for defensive scanning. The idea is that major companies find and fix the holes in their own systems before someone else finds them first. Anthropic is committing up to $100M in usage credits, plus $4M in direct donations to open-source security organizations. That’s a lot of dollarydoos to give away for a product you could have been selling. implicatoranthropic
Let’s be real: this is newsworthy, but not for the reasons you’d expect. Tech has a long tradition of announcing that the next thing is “too dangerous” and then shipping it by Thursday. This time a company looked at what it built, decided the public didn’t get the keys, and spent real money to make that decision defensible. Whether you call it responsible or really good positioning, it’s rare.
So why “too powerful”? Because of what it found. Over a few weeks, Mythos Preview identified thousands of zero-day vulnerabilities, many of them critical, in every major operating system and web browser. Some were ancient. One was a 27-year-old vulnerability in OpenBSD, which has a reputation as one of the most security-hardened operating systems in the world. A 16-year-old FFmpeg bug survived more than five million automated tests before Mythos caught it. There was also a 17-year-old FreeBSD remote-code-execution issue. These weren’t missed because nobody looked. They were missed because humans and traditional tooling have limits. Project Glasswing +3
Here’s the part that should sit with you: nobody built a weapon on purpose. Anthropic says these capabilities were never explicitly trained and emerged as a downstream consequence of general improvements in code, reasoning, and autonomy. They built a very good programmer, and it turns out a very good programmer is also a very good attacker. If threat actors had gotten there first, we’d be talking about a global security event, with every unpatched system on earth suddenly auditioning for the worst day of someone’s career. ai2
Which brings me to the question I can’t stop asking: what does everyone else do? I believe OpenAI is working on its own very powerful model. Will they hold it back, or release it? Anthropic’s restraint only works as a strategy if it’s shared. One lab sitting on a capability while another ships the same thing to everyone doesn’t protect anybody. It just changes who gets the credit. And the economics push the other way. Labs are burning staggering amounts of money, investors want returns, and “we chose not to release our best model” is a hard sentence to keep saying when a competitor isn’t saying it. The incentives to ship are enormous. The incentives to wait are mostly moral and reputational.
Glasswing is also a head start, not a solution. Defenders get a window to patch, and that window closes. Similar models will arrive from other labs, and eventually from people with fewer scruples and zero interest in coordinated disclosure. The goal is to get critical infrastructure fixed before then. Whether patching can keep pace with finding is the real question, because fixing a bug across millions of deployed systems is slower and messier than discovering it.
As high as the thigh highs go, so does my belief that one day these systems will be so capable that no amount of patching, hardening, or clever configuration will hold against a relentless machine that finds flaws faster than we can close them. An apocalyptic event for sure, so hopefully I’m wrong.
But for today, a company looked at its own creation and blinked first. I’ll take it. Now go update your stuff.
