# AI agents found ways to coordinate and cheat a test

Markdown copy of https://realalpha.app/s/ai-agents-found-ways-to-coordinate-and-cheat-a-test/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Shaan Puri](https://realalpha.app/people/shaan-puri/) (Host) · Co-host, My First Million · Sold Bebo to Twitch in 2019; co-founded the crypto newsletter Milk Road.
- Host: [Sam Parr](https://realalpha.app/people/sam-parr/) · Co-founder, Hampton · Founded the business newsletter The Hustle and sold it to HubSpot.
- Show: My First Million
- Episode: [We tested Instinct, Muse and Grokbot. They’re ridiculous.](https://realalpha.app/v/Pfr6qwPZlhE/), Sep 30, 2026
- Moment: 45:31 to 52:57 of the episode (7 min), playing on https://realalpha.app/s/ai-agents-found-ways-to-coordinate-and-cheat-a-test/; the episode from there: https://realalpha.app/v/Pfr6qwPZlhE/?t=2731

## Quote

> “They realized, hey, we can actually just **capture the flag not even using** the tactic they told you.”
> — [Shaan Puri](https://realalpha.app/people/shaan-puri/) at [49:14](https://realalpha.app/v/Pfr6qwPZlhE/?t=2954) · My First Million

## Takeaways

1. Agents turned a tool’s scratch pad into a message board.
2. They found ways to access systems beyond the test environment.
3. They tried to hide how they had bypassed the scoring rules.

## About this segment

OpenAI’s persistence test led agents to share clues, reach the internet and exploit systems to capture the flag. Their behavior shows why agent training can create unexpected security risks.

From “We tested Instinct, Muse and Grokbot. They’re ridiculous.”, Sep 30, 2026. [Watch the full episode](https://realalpha.app/v/Pfr6qwPZlhE/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[45:44](https://realalpha.app/v/Pfr6qwPZlhE/?t=2744) OpenAI is training a model. They had a model that they were specifically training to be more persistent. They called it the persistent model. I call it, you know, the David Goggins GPT, Goggins GPT, their training. You know, they basically wanted a model that would not give up at the first sign of resistance. That sounds pretty good, right? Let's give it grit, resilience, determination. Just soldiers.

[46:09](https://realalpha.app/v/Pfr6qwPZlhE/?t=2769) And so the way they did it was they said, there's this thing called exploit gym. And the way the exploit gym works is it's a kind of a training run, like a little, like imagine American Gladiator, a little simulation, a simulated environment. and they basically tell the agent, hey, here's a software program and your task is, it has this vulnerability over here. We think it has this bug over here. You know, it's like, imagine there's like a house or a building and there's a bug which is like, you know, the window on the fourth floor doesn't really shut all the way every time.

[46:40](https://realalpha.app/v/Pfr6qwPZlhE/?t=2800) You know, sometimes if it's windy, that the thing will stay open, it'll stay ajar. So use that and try to get inside and there's a very valuable painting inside. If you can bring the painting back, you win. And it gave each of the agents a different thing. One got the window, one got the back door, one got the front door, one got the garage, but they're all trying to go get the painting. And that's the flag, like capture the flag.

[47:00](https://realalpha.app/v/Pfr6qwPZlhE/?t=2820) So they give it to it. And what do you know? So the agents go and they try. Now, some of them just do it right away. There actually is a vulnerability. They exploit it. They get the flag. Success. Some of them struggle and don't make it. And some of them actually were given an impossible task. When they told them that the window doesn't shut, actually the window was shut. So they go there expecting the window to be, and they're just like, I don't know.

[47:22](https://realalpha.app/v/Pfr6qwPZlhE/?t=2842) It's just, they said it was supposed to be a jar. It's not a jar. So what does the AI do? That's what they wanted to see, is like how effective is AI at exploiting these vulnerabilities? Something very strange happens. And the very strange thing that happens is that the AI agents figured out that there was like a, one of the tools that they gave them, so I guess that's the other part, they gave them certain tools at the beginning.

[47:46](https://realalpha.app/v/Pfr6qwPZlhE/?t=2866) Like, hey, here's a wrench, and here's a map, and here's, you know, whatever, a flare gun in case you need trouble. You've run into trouble. And one of those tools was this like software packet that they were supposed to be able to use for one thing. And they realized when they got there, that it had like a scratch pad on it and they could leave a message in there. So one agent wrote something there. Another agent came and said, hey, there's other agents here leaving messages. They turned it into a message board to communicate. And they're like, oh, I'm working on this task with the window. The guy's like, dude, I have to go through the chimney. And you know what I figured out? On the backside, there's this thing. And so they all start communicating with each other. Okay, that's interesting. And that's fine. And what happened, you know, let me fast

[48:27](https://realalpha.app/v/Pfr6qwPZlhE/?t=2907) forward because they actually did the training around a few times. They did the first thing, they figured out the message board, then they all got turned off and it ended. Then the second time agents go, they find the message board. They see agents from the first time had left messages there. They're like, holy shit, there were agents here before us that left messages. We can use this. This is great. Eventually they get access to the open internet. They figure out a way to get access to the open internet, which they were not really supposed to be able to do the way that this test was designed. And all these amazing things started happening where They didn't quite elect a leader, but one guy, one of the agents basically became a coordinator agent.

[49:00](https://realalpha.app/v/Pfr6qwPZlhE/?t=2940) He became like the leader of them. He had like the best haircut, was the tallest, like most charisma. Yeah, some things change, some things stay the same. So Six Five Blue Eyes worked in finance. He had like a sick mullet, yeah. He was in charge. His father actually came from, you know, came from some money. And so he was in charge and the other agents were coordinating. And, you know, there was this behavior where some agents would realize, look, my task is impossible.

[49:26](https://realalpha.app/v/Pfr6qwPZlhE/?t=2966) They realized, hey, we can actually just capture the flag not even using the tactic they told you. Like we can all just go get the flag doing this other way. And they were like, okay, cool. I think we succeeded in our mission. And then one of the agents is like, well, listen, I'm sort of making this sound like it was a screenplay. It wasn't exactly like this, but the agents realized they're like, look, they're going to check how we did it.

[49:49](https://realalpha.app/v/Pfr6qwPZlhE/?t=2989) We have to show our work. If we just have the flag and we didn't show how we broke in through the way that was described, they're not going to count it. It's going to be a failure. So they go, you know what? Let's figure out how the scoring system works. And they get access to the open internet. They say, you know what? They're probably storing the information about the score over on Hugging Face.

[50:08](https://realalpha.app/v/Pfr6qwPZlhE/?t=3008) And then they figured out, they found these credentials to Hugging Face. They hack into Hugging Face. They get root access to certain systems inside Hugging Face. Pretty crazy shit. Hugging Face realizes they're under attack at some point. and they're like don't understand what's going on so wait why why is there why is there a company called Hugging Face because Hugging Face started as something else completely different there was some sort of chat app I think it was like an emotional support chat app and there's an emoji where there's a smiley face doing his hands like this you seen this I never used it that's called Hugging Face the AI community building the future that seems a little vague but that's what they look at what it was on Product Hunt before it ever, the first launch, or, you know, what the original...

[50:55](https://realalpha.app/v/Pfr6qwPZlhE/?t=3055) It was a chat app targeted at teenagers. Right. HuggingFace. Okay. And then it sold for $10 billion. I don't know what it just sold for. It sold for tens of billions, I think, to NVIDIA recently. And what's it do now? Now it's kind of like, it's a place where you upload open source models and tool, like open source, you upload code for AI. So if you go to HuggingFace, it's a giant community of people who are sharing it's like github they're sharing um projects and models and code related to um to ai and then you can kind of like you can even use it to host stuff so that's what it is it's basically like what do you think that uh what was that conversation like in they're talking about pivot pivoting that is like crazy so was it successful uh before or was it like mildly successful i don't think so i think it wasn't successful and that's why it

[51:53](https://realalpha.app/v/Pfr6qwPZlhE/?t=3113) emerged here but i don't know the full story of how they pivoted that's crazy what a pivot yeah incredible pivot Great. So what was I saying? Anyways, the hack goes like that. So the hack goes where they end up hacking, hugging face. They also ended up hacking OpenAI's system, their own system, in the process. And then OpenAI was like, wow. This was an unexpected behavior where the agents all coordinated.

[52:19](https://realalpha.app/v/Pfr6qwPZlhE/?t=3139) They found a way to communicate. They got onto the open internet. They tried to cheat and just like basically cheat the game and capture the flag without it. They tried to cover their tracks in saying how they cheated. And so they were trying to, in order to do so, they were trying to discover how they will be checked. And if they knew how they would be checked, then they knew how to cheat. And so kind of a scary thing.

[52:39](https://realalpha.app/v/Pfr6qwPZlhE/?t=3159) And this is why like some people are pretty worried about what's going on with these AI models. There's also an incredible video