Cal Newport

This has been a huge concern.

And it's a real concern, by the way, right?

Is that partially what...

slows down slightly cracking, right, the breaking into systems, is the fact that it's annoying and hard.

And LLMs have made it easier.

GPT-4 was good at finding exploits, right?

And this was a big deal.

They were like, GPT-3, 5 wasn't great at it, GPT-4 is.

And then as we got the more recent models, they've been much better at writing code to exploit them because we had better agents for it and they're more, they're

They're better able to produce multi-step software goals, and so they can better build software to exploit them.

This is a real issue, but it's not new with Mythos, right?

But Mythos was presented as if some Rubicon had been passed.

But there was a couple things I noticed right off the bat.

One, they made the mistake of listing

a bunch of the exploits that they, vulnerabilities they had found to try to brag.

Look at this thing in FreeBSD.

Look at this thing in FFPG or whatever.

Like they showed all these exploits they found.

They didn't count on a lot of security researchers said, well, wait a second.

Why don't I get like a much smaller, cheaper model aiming at that same source code and say, can you find any vulnerabilities?