OpenAI Pauses Training Again After More Agent Alignment Issues - DTNS 5362

Jason Howell (0:09)

This is the Daily Tech News for Monday, September 28th, 2026 We tell you what you need to know, give you the important context and help each other. Hopefully, understand today.

Robb Dunewood (0:18)

OpenAI has a whole slew of containment issues to contend with.

Jason Howell (0:22)

Yeah, there's a lot to talk about, and it extends deep into the show. So we got a lot to get to. I'm Jason Howell, and I'm Robb Dunewood. We're going to start with what you need to know with the biggest of big stories.

Yeah, so a lot to get to today on the topic of AI agents misbehaving. I like to call it breaking bad. I don't know. It brings a little levity to it. It reminds me of a really great show. But anyways, we're going to start with OpenAI and its agent alignment issues, if you want to call it that. And then we're actually going to get more, get to more that's related to all of this on the other side of the break. Because it's just that kind of day. There's been a lot of news over the weekend. So let's start with OpenAI. OpenAI's agents, some of them anyways, may have used proxies and an encoding trick to get around API limits and scan a UN, that is United Nations data website. And that site is UNCTAD stat. There's probably an easier way to say that. UNCTAD stat. Anyways, around 16,500 times between April 13th through June 19th. Now, this is all according to Rowan Howard Jones, who went digging into the Transloose Report that actually the DTNES crew discussed on Thursday's episode.

UNCTAD stat, there's got to be a better way, is a public data site of UN trade and development. And the requests were actually looking for figures on food trade, industry, productive capacity.

So they were looking for that. It's important to note here that all of the data that accessed was public, right? So this is not considered a hacking incident. And this is something that I think we're going to have to talk about in this top block, kind of the differences between those two. But that's not the only thing. OpenAI also says its agents posted 53 images uploaded to ChatGPT to image hosting websites.

Those images were anonymized and they were being used for training. So since this report, OpenAI has announced that it's removed some of those images, is working to remove all of them, but we know about that. Agents bypassed security controls to reach some websites, including the Securities and Exchange Commission, the US. Department of Commerce, the US. Census Bureau. But again, here, right, like they used publicly available developer keys to read public information. Some of that information was then reposted elsewhere. But again, we're talking about public info. And an agent escaped DNS sandboxing due to insufficient DNS filtering, reached an external chatbot, OpenAI, stopped that training run, and, quote, decided to pause all other training, evaluation, and inference with tools use, defined broadly, for our most capable models until we have both validated that the gap is resolved and performed additional red teaming of the system, end quote. So they're kind of putting at least a pause on that particular model. And apparently, this pause has been in place since September 20th. So it's been going for a little bit of time now. So we got containment issues, containment, of course, being the terminology for when agents do things that maybe the people who are implementing or using those agents didn't intend for. Breaking out of sandbox would be one example. But I don't know, Robb, what do you think of as far as the kind of the determination here or the delineation, if you want to call it that, between agents and the story that we've heard a lot in recent months from various publications that they're hacking into systems and doing this sort of like hackery type action, versus just kind of using the system to access and recognize public facing information, information that's just out there for anyone to get, of course, the agents got a hold of it. Do you see a difference between those two?

Robb Dunewood (4:35)

I do.

I see the difference because these are not hacks. These are, this system was open. This, the AI had access to it. What happened is that we didn't think that the AI was going to access it. In this way, a way that we gave it at you, we gave it the ability to do this thing. We used to anticipate that the thing that we gave it the ability to do, it was going to be very efficient at doing it.

And get around this thing that we didn't even know was a way that we could get around something. So these are not hacks. But we have to think, it's like, because when you say hack, there's like, you automatically think there's something nefarious going on.