Cal Newport

because it was absolutely brilliant what they did there.

I've spent a lot of time on it.

It's complicated because, again, you can't really trust the system.

The system cards are just gonzo that Anthropic puts out, and it's not publicly available.

But there were, I think, a few very telling things.

So there's two features they say Mythos has.

One is finding vulnerabilities in source code, and two is writing programs to exploit them.

It's first really important that people understand

this has been something that people have been doing with LLMs since the beginning of publicly available LLMs, right?

Not only is there nothing new about that, but I found, I put this on my podcast, almost word for word from the Anthropic System Card, they said in the Opus 4.6, rather, Systems Card, right?

A publicly available model that's already been out for many months.

Almost word for word for what they said about mythos, except for no coverage of it and no fear.

They said, we have found

500 zero-day vulnerabilities, including some that had been existing for decades without having been discovered.

That is what they said about what Opus 4-6 could do.

For Mythos, they said the same thing.

They just replaced the word 500 with thousands.

But when Opus 4.6 came out, there was no, oh my God, they have found many hundreds of zero-day exploits, many of which have been around for decades, because they didn't push that marketing button.

No one particularly cared about it.

I went back to my podcast and showed multiple papers.