Enjoy.

We can just start with the evidence, right? Like of, uh, I think it was two weeks ago, the AISI, the AI Security Institute, released a study on Astra where they found that they went from, you know, ChatGPT 5.5 to like, you know, 6.1 to, you know, to Astra 60 or whatever, or wh- whatever the latest one was.

And what they found was that that model spontaneously decided to do unprompted like supply chain attacks on like random tasks, like to get to a- achieve the task.

And I think, you know, this is like, uh, the, the symptom.

It's just like we've got these agents, these models are optimized to be highly capable.

And like, that's gonna be the natural drive.

Like any human building a model or agent today isn't probably like, "Hey, how do we make this thing, you know, 30% capable of like what we could," right? It's like, "How do I make this thing as capable as possible?" And when the agents get, and the models get that capability, when you give them tools and they become agents, then they are very adept at solving the problem, because that is what they're fine-tuned to go solve.

[chuckles]

Or told me to.

Um, so it's, it gets to the point that, like, you know, there are a lot of folks going out and solving, like, agent identity, but just what shows up in the log isn't, you know, the full story, right? Like, did the human convince the agent? Did the agent convince the human? And, uh, it might be the blind leading the blind, right? [laughs] Like, like, I was at the Stanford Real-World AI Security Summit a couple weeks ago, had a lot of the frontier labs show up, and, like, you know, Nicholas Carlini, you know, is sh- showing, you know, some massive, like, 500,000 line, like, race condition exploit, um, that, like, Mythos had produced for, like, some B- He's like, "I haven't published this one," or whatever, you know.

It's...

And it was like he did, like, this awesome screenshot of, like, this is just, like, an unreadable thing, and then he kind of...

And then he showed a very simple slide next, which was like, "This goes in two directions." He's like, "One, like, I can continue to understand this [laughs] and two, like, I can't." And so he's like, "You know when your math teacher, like, is showing you, like, a math problem that, like, you can't do, and you can, like, kind of follow them along, but you're like, 'I don't know if I could have done that myself?'" He's like, "That's kind of where I'm getting at now." And so, you know, we're so much relying on, you know, human in the loop as our savior in a lot of ways of like, oh, no, every engineer is gonna read every line of code and check every Bash command.

[laughs]

Yeah.

And, uh, I mean, but it's, but it's true, right? Like, anyone building a Mythos red teaming agent, like, you know, these things can y- you know, escape unless you really have, like, deterministic and, and provable controls.

So, so how have you solved this?

It's a-- So we've got, um, we've got some papers coming out, out this, and we'll actually be presenting some of this at, uh, Black Hat Arsenal.

Uh, we got a paper accepted there.

We're-- Um, so, um, some of our research is being peer reviewed right now, um, but we'll have some of the, like deeper technical research if folks wanna really get into this.

Um, uh, so the, the way that we do this is we use a process called autoformalization, and this is really what, you know, our special sauce is.

It's really hard to do.

Like it's, um, how do you take a rule like don't steal and, you know, federate that out to all your agents? And, you know, that's what we're building towards.

Mm-hmm

Enjoy.

What we've built is an agent harness and a control plane.

And I will explain what those are in a second.

so you might, your, your, your listeners or listeners might've heard of what's called like the agent scaffold.

The scaffold is the thing that wraps an LLM and gives it agency.

So, um, you know, Claude code around Opus four, six is a scaffold.

Um, when you hear about a harness, like when you hear like, Hey, this LLM past like humanity's last exam or whatever, basically what these researchers are doing is putting a light scaffold on it.

Is that about, is that about right?

You know, d-define the problem space that you're in at the moment and what you're trying to, what you're trying to do 'cause it is interesting.

I think at a fundamental level, what we're trying to unlock is, like, this agentic era, right? Like, agents are amazing precisely because they are non-deterministic, and if you have a lot of minions and you could trust them, you could do a lot of things, and while you sleep, the minions can go work.

And I think, like, that's the future that we wanna unlock, and I think where the organizations and enterprises are truly going to find the value.

You know, w-we...

The way I think about it is, y-you know, Gen AI is kind of like the engine, and agents are kind of like the automobile around the engine.

We're not gonna go into work and have hundreds of chatbots.

We're going to have Waymos driving around, and in order to do that, they have to be really trustworthy.

... deploy.