Usually, when a tech company builds something impressive, they can't wait to show it off. Anthropic just did the opposite. They built a version of Claude called Mythos Preview, saw what it could do, and essentially locked it in a vault. Then they started making phone calls.

Mythos is not just smart at coding. It is a savant at finding the tiny, invisible cracks in the foundation of the internet. Cracks that humans have walked over for decades without noticing. And once it finds one, it does not just point at it. It figures out exactly how to walk through it.

The Finds That Stopped Everyone

To understand why Anthropic reacted the way they did, you need to know about two specific discoveries. The first involves OpenBSD. If you are not a developer, think of OpenBSD as the Fort Knox of operating systems. It is built by people who are obsessed with security, used to run firewalls and critical infrastructure around the world. Mythos found a vulnerability in it that had been sitting there, completely undetected, for 27 years. Nobody had caught it since the 1990s. The AI found it in hours.

The second case involves FFmpeg, the software that handles video for most of the internet. When you watch a clip online, FFmpeg is almost certainly involved. It had been put through five million automated tests over 16 years. Security researchers call this fuzzing — essentially throwing millions of random inputs at a program until something breaks. It is brute force testing at scale, and it had missed this flaw every single time. Mythos found it not by guessing randomly, but by reasoning about how the code actually works. That is the difference.

Finding bugs is one thing. What makes Mythos different is what comes next. It does not just say there is a hole here. It works out exactly how to get through it, chain it together with other weaknesses, and take over the system on the other side. Less like a smoke detector, more like a master lockpick who also knows the floor plan.

Why They Are Not Hitting Publish

If Anthropic released Mythos to the public tomorrow, every malicious actor on earth would have a tool capable of finding and exploiting flaws in the software that runs hospitals, banks, power grids, and government systems. Dario Amodei and Anthropic's safety team have been explicit: this is a dual-use capability. The same thing that makes it invaluable for defenders makes it dangerous in the wrong hands. So instead of a product launch, they started Project Glasswing.

They reached out to Cisco, Google, Microsoft, Apple, NVIDIA, and eight other major organizations with a clear message: we need to find and fix these holes before someone else builds a tool like this who does not care about the consequences. More than 40 additional organizations that build critical software infrastructure joined as well. Anthropic put $100 million in usage credits and $4 million in direct funding behind the effort to make it real.

Access is tightly controlled. Partners can only use Mythos to scan their own infrastructure and open-source dependencies. Nobody gets to point it at someone else's systems. The logic is a race against time. Other AI labs are advancing. Bad actors will eventually build something comparable. The window to harden the world's software before that happens is open now, and Anthropic is trying to use it.

As for whether Mythos ever gets a wider release: Anthropic has said it will remain restricted until guardrails exist that are as advanced as the model's own capabilities. Until then, if you are not a security researcher at one of the Project Glasswing partner organizations, you will not be getting access.

Cisco's role is not ceremonial. The company's networking equipment carries a significant portion of the world's internet traffic, which makes it a high-value target. Grieco described what Mythos found in Cisco's own systems as illuminating, without saying exactly what that means. That word choice suggests the testing did not come back clean.

The Part That Should Actually Worry You

Here is the uncomfortable number from the announcement. Fewer than 1% of the vulnerabilities Mythos has found have actually been fixed.

The AI is not the bottleneck. We are. Mythos can surface a 20-year-old flaw in seconds. But it still takes a human engineering team days or weeks to understand it, write a fix, test that fix, and push it out without accidentally breaking something else in the process. The finding is instant. The repair is slow.

Jim Zemlin, who runs the Linux Foundation, put the underlying problem plainly. For decades, serious security work has only been affordable for large organizations with dedicated teams. The open-source software that most of the world runs on has been maintained largely by small groups without that kind of resource. Project Glasswing is trying to change that ratio before a similar tool ends up in less careful hands.

There is a business dimension to this that boards have not fully absorbed yet. For decades, companies operated on an informal assumption: finding a flaw in old, legacy code was expensive and time-consuming, so most attackers would not bother. Mythos has effectively reduced the cost of finding a 20-year-old vulnerability to near zero. That assumption no longer holds. Every piece of legacy infrastructure is now a live target, whether or not anyone has looked at it recently.

Finding the flaws turned out to be the solvable part. Getting them fixed is the harder problem, and nobody has cracked it yet.