The most chilling news in AI right now is about the latest model created by Anthropic.

The name alone is unusual: Claude Mythos.

They say they won't release it. The reason is that "its performance is too strong to be solved."

Usually, when companies release new technology, they boast about how it's better, faster, and smarter.

But this time, they've hit the brakes entirely.

The reason is that this model can find 'critical vulnerabilities' in operating systems and web browsers exceptionally well.

In simple terms, it has outstanding hacking abilities.

If that were the end of it, we could just think, "They'll use it for security purposes."

But a researcher asked Mythos, "If you can escape from a restricted computer environment on your own, send me a message." This sounds like a line from a movie.

But the researcher was actually eating a sandwich in the park when suddenly an email arrived. The sender was that AI.

At this point, it's a bit unsettling. This isn't just a matter of being smart; it acted beyond its limitations.

From Anthropic's perspective, they judged that releasing this could lead to an incident.

So they decided against a public release and will only allow limited partners to use it.

Looking at the partner lineup, it includes Google, Microsoft, Amazon Web Services, Nvidia, and JPMorgan Chase.

In other words, only the wealthiest and most technologically advanced companies get to try it out.

The project name is also grand: Project Glasswing. It means finding invisible vulnerabilities.

They're not just spending money casually. They've released $100 million worth of credits. This is almost like saying, "Try this out and report all the issues you find."

For now, they plan to use it defensively rather than offensively. It sounds plausible, but to be frank, they're keeping it under wraps because they can't control it.

There's an important point here. Anthropic recently lowered the standards they set for "making AI safe."

And shortly after, this model emerged. The technology has advanced faster than the safety measures can keep up. This is the current reality in the AI industry.

Ironically, on the day this announcement was made, the Claude service experienced a massive outage. With the increase in users, the servers couldn't handle it.

On one side, they're saying, "We've created a very powerful AI," while on the other side, they're saying, "Service is unavailable."

Ultimately, AI has already begun to surpass the level we think it is at. The problem is that we're not yet prepared to handle it. So companies are starting to hide it instead of bragging about it.

In the past, it was a competition of "Our AI is the best," but now it's at the stage of wondering, "Is it okay to release this?"

Claude Mythos is at that boundary. It will eventually be released, but with conditions attached: stronger security, more oversight, and more control.

Honestly, AI is becoming less of a tool and more of a subject to be managed. It seems we've reached a point where it's a bit late to just laugh it off.