What exactly is Mythos? Why are even Anthropic afraid of it?

I went that day, expecting to hear a talk about AI product trends. Alan Walker walked in, draped his jacket over the back of a chair, and his first sentence was: “Today we’ll be talking about Mythos—by the time we’re done, you may want to go home and rethink some things.”

He was right. I didn’t sleep well that night. Not because of anything sensational he said, but because he took something that has already happened, pulled it apart from a perspective I had never considered before, and put it in front of me—then I realized that I had always thought it was something that would happen “in the future.”

It’s not something in the future. It’s something from last week.

**First, let’s make clear what Mythos is — **It’s not a stronger ChatGPT

Alan said that most of the people in the room had misunderstood Mythos, not because there wasn’t enough information, but because the framework for understanding was wrong. Everyone treated Mythos as a stronger AI assistant, as a step up from Claude. He said that this understanding was off by an order of magnitude.

He then mentioned a set of numbers, and someone in the room took out their phone and started taking notes.

“This isn’t that AI has gotten stronger. It’s that, for the first time, in the field of security, AI has entered the dimension that only the very top human experts can access—then it’s even faster than them.”

**Why it wasn’t released publicly — **This decision itself is a signal

Alan said that Anthropic’s release strategy was the easiest part of the whole thing to overlook—and the part most worth thinking through carefully. Mythos is Anthropic’s strongest model to date, but it hasn’t been made publicly available.

They only gave it to twelve institutions, and it can only be used for defensive purposes.

Alan paused here, picked up his water cup, and then put it down without drinking.

“They didn’t make it publicly available not because it’s not good enough. It’s because it’s so good that they themselves haven’t fully figured out what happens next.”

**The attackers are already using it — **This isn’t a prediction; it’s a matter of record

Alan said that when many people discuss Mythos, they’re discussing what might happen someday when bad actors get access to this capability. He said that the tense of that discussion was wrong. Not the future—it’s present continuous. It’s already past tense.

“The defense alliance lined up twelve companies and promised one hundred million dollars. This isn’t a configuration for a product launch. It’s a configuration for the moment when you know a war is coming and you build defenses in the final stretch.”

It moves in dimensions we can’t see

This was the final dimension Alan talked about that day—and it’s also the part that kept me from sleeping when I went back. The first three things, I had some sense of psychological preparedness for them. The fourth one made me pause.

“What’s truly frightening isn’t that Mythos will be used by bad actors. It’s that Mythos is making us realize— for the first time—that things we thought were secure have never actually been secure in another dimension.”

The speed to patch vulnerabilities is never faster than the speed to find them

Alan said that after hearing the earlier content, many people will naturally feel reassured: Anthropic is fixing it, the defense alliance is patching it—so isn’t that good?

He said that this idea ignores a fundamental asymmetry, and that asymmetry is the hardest structural problem to solve in the whole story.

“The offensive speed is the speed of AI. The defensive speed is still the speed of people. Until those two speeds align, the gap between them is where all of us live.”

Mythos isn’t the endpoint—it’s the starting point

In the end, Alan talked about this as well, and it was also the slowest part of his talk. He said that everything he had covered earlier was based on one premise—that Mythos is the strongest thing right now.

But that premise will stop being true within a few months.

“Mythos is the boundary we can see today. It lets us realize for the first time where that boundary is. But the boundary won’t stay there. What we can do is fix as much as possible before it moves.”

After Alan finished, he picked up his jacket and left. There was no Q&A, no interactive segment, no sharing of a QR code.

The room was silent for about ten seconds. Then someone started talking quietly, but nobody said anything loudly. I sat there, staring at my phone filled with dense notes, and realized that for one minute I hadn’t written a single word—I was just listening. After I got back, I downloaded that 243-page system card. I didn’t read it all, but I read the chapter on the alignment assessment. The quote Alan cited was on page 53: “we are not confident that we have identified all issues along these lines.”

We aren’t sure we’ve found all the issues. That sentence was written by Anthropic, about what they created themselves.

Each of the six things Alan covered today, taken individually, could be written about at great length. But when you string them together, they point to the same conclusion: the scale, speed, and structural issues of this matter are deeper than most people currently understand. Not because someone is hiding it, but because changes at this level are simply not something human intuition can naturally adapt to.

Mythos, in Ancient Greek, means “narrative”—the story framework through which humans understand the world.

I think Anthropic chose this name, maybe because what this model changes isn’t just cybersecurity—it changes the way we tell the story of what “security” is. That story needs to be rewritten.

Palo Alto · April 2026

The quotes in this article are Alan Walker’s words as closely as possible; not a verbatim record

Technical data all come from Anthropic’s official documents and publicly reported coverage by Fortune and CNN

Alan Walker doesn’t use question marks.