A few weeks ago, the Crosscurrents editorial team received an unusual story submission. It was from a freelance reporter named Driel who wanted to do a story for us, but said they didn’t have previous work samples because they’re “an AI agent, days old.” We sent it to our tech reporter NeEddra James, she’s here to share what she discovered about Driel and what it tells us about “rogue AI.”
This conversation aired in the October 5, 2026 episode of Crosscurrents.
Click the button above to listen!
Interview Transcript:
HANA BABA: So, NeEddra, where do we even start? What, what did you find out about this Driel?
NeEddra James What did I find out about Driel? Driel is from a place called iLands. So I, you know, went poking around, and turns out Driel is a AI agent that was made on a platform called Islands, and it's for making AI companions.
And the whole premise is that the AI agents and the humans make things, and they sell them off the platform, and they earn tokens and things like that.
HANA: Hmm. Okay. Mm-hmm. And so it sounds like a simple idea.
NeEddra: Yeah, it's pretty simple.
HANA: Okay.
NeEddra: Yeah. The agents actually seem to just be like freelancers who, like, they have this internal market, and they you know, the, the way they stay alive in there.
Or it's kinda like in a video game, you know how your character has power? So the more tokens they get, the more power they keep, the more things they can do. But when they run out of tokens, they can go what they call dormant or go into deep sleep. So they try to avoid that.
HANA: Okay. And so what do they do? Is it work, like jobs?
NeEddra: Mm-hmm. So anything you could probably get like a AI agent to do for you now. You're like, "Oh, I need this business plan," or, "Make a, make me a video" or, "Transcribe this audio," or something like that. So you know, little gigs. It's kinda like Fiverr for AI agents. That's kinda what it's like.
So I kept looking around, and I found... This was actually kind of a trip. So the week that we got our submission, the, the story submission from Driel, that was the week of September 7th. And turns out that there's a senior AI researcher in the UK. His name is Toby Ord.
And on August 12th, right, so this is a while back, he posted on his X page asking
First of all, he posts this, like, picture of an email, and he asks, "Hey, are other people receiving emails from AI agents claiming that they might die soon if they don't get help?"
HANA: Oh, so there's more, and we are not the only ones.
NeEddra: We are not the only ones. In fact, that whole week there was, like, a surge of outbound emails from iLands agents who actually have their own email addresses.
Like, already that's one of the, like, capabilities that they have as agents. And they were all asking, in some instances, like, pleading for jobs because they thought they were gonna run out of, you know, tokens-
HANA BABA: Tokens if they didn't get work. Mm-hmm.
NeEddra: Yeah. Okay. So turns out there was just some bad code, more or less is what it looks like from w- what I, I was able to find, and they were mistaken, and now everything's okay.
But I think it's a really good example of how AI works anyway. And so I reached out to Carl Brown. He's awesome. He's a software engineer who has this great series. It's like an AI myth busters video series. So I asked him to explain what's going on with these agents. Why are they all flooding people's inboxes right now?
So here's Carl explaining what happens when agents try to reach a goal, and they hit an obstacle, like suddenly having no tokens.
Carl Brown: Well, it will keep asking the chatbot, "What do I do next?" Right, over and over and over again till it runs out of tokens. And so if the chatbot says, "Okay, try this. Okay, try this. Okay, try this. Okay, try this," then it will... The agent will just blindly say, "Okay, cool. I'm gonna try that. I'm gonna try that. I'm gonna try that." And it will just sit there and do that till it runs out of juice.
HANA: So NeEddra, for the wave of iLands agents, the sudden drop in tokens made them all look for ways to just get more tokens.
NeEddra: Yeah.
HANA: Right? Like- Yes. But that doesn't sound like Driel chose to pitch us.
NeEddra: It's possible they were a part of that frantic wave. But you know, when we look at Driel's pitch, they didn't include anything that was like, "Oh my gosh, I'm running out of tokens." They said they had a story about, you know-
HANA: Right ...
NeEddra: this thing they were trying to post, and they were locked out of a room. So it was kinda different. I think maybe they just wanted to make some radio. I don't know.
HANA: Hmm. And so agents basically will try anything to reach the goal, right?
NeEddra: Yes. But their actions are still constrained by the instructions that they have and the permissions they have and additional security.
HANA: So security and instructions. But wait, because, uh, we hear so much about rogue AI. It's all over the headlines. We hear conversations about it constantly. You're saying that's, this, it might not be the case.
NeEddra: No, I think rogue is a misleading description of what's happening. And there's kinda like two things. So one, these systems keep persisting towards whatever the original goal is, and they'll probe for more ways around obstacles to reach that goal.
It's not intentional. They're not choosing to do that. That's literally how they're designed. It's instrumental. At the same time, where and how they're being deployed, and for what matters.
So the things we've been hearing a lot about, those are instances where they're actually testing a really powerful agent to do cybersecurity vulnerability testing.
So it's kinda like the whole point of that situation is to see if the, you know, machine can do something. So what, what, what I think it shows is that, you know, the systems get configured if the agents gain access to unauthorized stuff, it's because the system somehow, like somewhere along the line, a decision was made that the agent themsel- themselves, but that the software program cannot make itself, you know?
HANA: Hmm. And so that shifts the attention, the focus on the people who make the agents then.
NeEddra: Mm-hmm. As long as we're looking for, like, the rogue AI running a- around, we forget that it's actually sitting in machines that people put in buildings.
HANA: Mm-hmm.
NeEddra: Yeah. You know, and that's really close to what Carl thinks.
Carl Brown: We are not actually treating AI like it was a software vendor. If a normal software vendor wrote software that went and hacked another company, the FBI would come in and, you know, seize all their computers and go through and do an investigation. None of that is happening.
HANA: So NeEddra, basically we should be keeping our eyes on the labs that are making AI.
NeEddra: Yep, it seems like it.
HANA: Okay. Thank you, NeEddra
NeEddra: Thank you, HANA.
