AI Agents Continue to Escape Their Sandboxes

Brian (0:00)

Hey, what's going on, everybody? Welcome to The Daily AI Show. Today is September 28th, 2026 And with me today is Andy to kick off the show. Hey, Andy, how are you?

Andy (0:09)

Good, thank you. Good morning.

Brian (0:11)

Good morning.

It's a fast start to a Monday. I think you've got your hands full on your property, and I'm doing stuff on mine. So it's, but we're here. We're ready to go. We're ready to talk.

Andy (0:24)

But AI never stops. So we have to jump in and talk about it.

Brian (0:30)

Yeah, I was saying before the show, I have a few things that I know that have been going on. You said you have some new stories. We'll definitely jump into those first. But the big thing I see everywhere is more AI ages, or not that they went rogue yesterday, but that we're uncovering now, I guess potentially thousands of different incidents.

Andy (0:51)

Tens of thousands of incidents at OpenAI alone.

And many of them are smaller. And to a large degree, the news is that as yet, there's no serious or significant damage. These attacks, as they're phrased, are their agents gaining access to and actually doing some things that they shouldn't be doing in there, but it's not with any kind of overarching design to do damage. So nonetheless, they should be in there. Yeah, they shouldn't be in there, right? So that's the that's the news.

Brian (1:28)

Yeah, yeah, yeah. I mean, the little bit I have really dug into, I haven't dug into it, I shouldn't say, but the bit I have dug into, goes back to the same sort of story is two things, or two takeaways, I guess. I'm going to clear my brain here on a Monday morning. Two things, one, that the OpenAI is the Anthropics of the World, can't do the research fast enough to catch it. I got sort of that impression. I'm sure I'm broadly stroking this, but that was one of it, was that even moving full tilt to try to catch things. In one case, Elena Isanoff on TikTok who does wonderful stuff. I talked about it a lot of times. Amanda, she did a great job of breaking things down. She talked about this and she said in one case, the human in the loop caught it within 15 minutes. Fifteen minutes when AI is moving as fast as AI is, but a thousand sub-agents is already a lie. Then the fail switch broke and they couldn't actually stop it for two and a half hours.

She points this out and you got a lot of issues here. You have AI agents that are trying to complete a task and finding every which way, and I think she's the one. I want to make sure I'm giving her credit, because I'm pretty sure she's the one where I heard this. But her take on it was it's not like these internet loopholes and errors and backdoors don't exist. The difference is you now have AI that can go find all of it almost immediately and organize around it as if you would have some super crime that could do this, but it's just beyond human terms.

At least the way I understood it, maybe Andy, if you heard differently. To your point, it's not that we have AI agents creating backdoors or quote unquote hacking. They're using pathways that exist in finding and exploiting those to achieve a certain goal. So even when it's been said to like, you shouldn't have any outside contact with the internet or whatever, they wrote code within images or something and had found a screenshot tool that allowed the agents, and I know I'm humanizing this, but the AI agents were able to use that code somehow in order to see the screenshot to figure out what they needed. It's like it really, really reads like fiction. It really reads like fiction and it's not, and that's the hardest part to wrap our heads around these days. So anyway, that was where my brain on it. What was your take on it, Andy?

Andy (4:07)

Well, so a lot of the containment that's required and failing at the moment, containing the most advanced models that are in research labs right now, that depends on a system called a sandbox, which is creating an environment that the agent can operate and be tested in that doesn't allow it to go out and do these things on the open Internet. But when you're testing agents, one of the key skills and values of an agent is that it can go out on the Internet and do things for you. So that part of it has to be monitored very carefully. Well, in the big argument around whether we should slow down or not, Jensen Wong has taken the position that most of these problems, if not all of these problems, can be solved through engineering.