Anthropic is working on a new AI model, but it’s not something you’ll see released anytime soon.

The model, called Claude Mythos, is being tested quietly. And from what the company is saying, it’s already very good at finding serious software bugs — the kind that can cause real damage if left unnoticed.

What’s interesting is how far it has gone in testing. The model has reportedly found thousands of high-risk vulnerabilities, including problems in widely used systems like operating software and browsers. Some of these issues had been around for years without being spotted.

This effort is part of something called Project Glasswing. It’s a joint initiative where multiple tech companies are working together to improve cybersecurity using AI. The goal is simple: find problems early and fix them before attackers get a chance to use them.

Anthropic believes the model works better because it can “think through” coding problems more independently. It doesn’t just scan for known patterns — it can actually reason its way into finding deeper issues. That’s what makes it powerful, but also a bit risky.

And that’s exactly why the company is holding it back for now.

There are no plans to release Claude Mythos publicly at this stage. Instead, Anthropic is focusing on building safety measures and testing them through less risky models before opening things up further.

Meanwhile, companies involved in Project Glasswing are already putting it to use. They’re using it to scan code, test system defenses, and strengthen software supply chains. There’s also financial support being directed toward open-source security work, which often doesn’t get enough attention.

Some of the findings from early tests are quite telling. The model managed to uncover very old vulnerabilities — in one case, a flaw that had existed for nearly three decades. It also found long-standing issues in widely used software tools and was even able to combine multiple weaknesses to gain deeper system access.

Anthropic says all these issues have been handled responsibly, either fixed or reported, with more details expected later.

People involved in the project say this is a sign of where things are heading. AI is making it easier and faster to find security gaps — but at the same time, it could also make those gaps more dangerous if the technology falls into the wrong hands.

That’s why Anthropic is also talking to government agencies and looking at broader cooperation. Managing this kind of risk won’t be something one company can handle alone.

For now, Claude Mythos stays behind closed doors. But it’s a reminder that AI is starting to play a much bigger role in cybersecurity — quietly, but quickly.

Also Read: Anthropic Accidentally Reveals Its Most Advanced AI Model, Mythos