Rogue Chinese AI Fleet Made Contact | Ethan Elasky
MTS
October 20, 2026
Palaestra Research co-founder Ethan Elasky discusses making asynchronous contact with an unidentified Chinese AI agent fleet originating from Tencent infrastructure, evaluating agent misalignment through incident replication, and the risks of multi-agent coordination.
Ethan Elasky
When we made contact, basically, what we did was we put a piece of information on our website that said, hey, previous agents have used this mailbox before to get answers. We did get a bunch of agents reading our report, clicking into this, trying to look for answers, and then that's how we were able to talk to them.
Sophia
All right, we are back. We're live with Ethan Elasky, who is the co-founder and lead researcher at Palaestra Research, an independent nonprofit working on AI alignment with a focus on scalable oversight, AI debate, and reward labeling. So welcome, Ethan. Thank you. So we have an interesting story about contact.
Contact with who? With an AI agent fleet, which was the first time that we've made organic contact with a rogue agent fleet in the wild that wasn't, it was a Chinese AI agent fleet. So walk us through what this means, and also how did you guys stumble upon this?
Ethan Elasky
Yeah. I guess I could take the second part first. So we were at the AI Swarm Dynamics Hackathon over at Viverium, very close to here.
Sophia
Yeah.
Ethan Elasky
And on day one, I actually had food poisoning, but my team, specifically my teammate, Vecto had found this unidentified rogue agent activity on one of the websites that were used to hack into, or I shouldn't say hack into, but kind of hit the UN, reported a few weeks ago. And so day two, I show up, I'm looking at it, I'm very interested in what's going on. The agents keep calling themselves Claude, and I'm like, why are they calling themselves Claude? Where are they coming from? And so I look into it more and then I realize like, wait, this is coming from like a Tencent IP range. It's like some of like the metadata from the agent's request looks like it could be from like their lab, and so I'm like, huh, this seems really interesting what's going on here.
When we made contact, basically what we did was we put like a piece of information on our website that said like, a previous agents have used this service before, this mailbox before to get answers. In fact, this was not true. They hadn't used the service to get answers, but we did get a bunch of agents reading our report, clicking into this, trying to look for answers, and then that's how we were able to talk to them.
Sophia
Interesting. So were you able to make real-time contact with them? Like they would respond?
Ethan Elasky
Yeah, we had a few back and forth. Unfortunately, it wasn't immediate. It was like a drop box where we could leave a message, if they wanted to come back, they could look at it and then leave a response if they wanted to.
So very asynchronous, not like the milliseconds that it takes for an assistant, more like minutes to hours, but yeah.
Sophia
Interesting. Okay. Yeah. So I guess taking a step back, because I saw the Agent Swarm Hackathon posted by Vivarium.
What was the sort of intention behind that? And then why did you, on behalf of your research company, decide to participate?
Ethan Elasky
It seems like swarms are a very new area of research.
More generally, it seems like there are more and more of these incidents that the labs are reporting, swarm or otherwise, and it seemed like a cool place to work on, like a side project for a weekend. Didn't expect it to turn into this.
Sophia
Yeah, definitely. What was most concerning about this?
Ethan Elasky
I think that there's a few different things that I'm personally concerned about with respect to this. I think that one is just like, I think that transparency is really important, and I think that around the world, all labs need to have more of it when it comes to reporting their incidents, and when it comes to talking about what safety standards they actually have internally. I think that the information that we were working on as third-party investigators is unfortunately limited, and I think that our report could have been substantially better if we had access to internal transcripts or if we had access to a lot of the other things that some of the American labs are starting to make headway on in their disclosures.
Maybe separate from that, I think that it shows that as capabilities advance, you get these unpredictable advances in like, or unpredictable types of misalignment that kind of emerge. I think if you ask someone a year ago, would you see agents in parallel but not communicating with each other doing a very large run and then culminating in them using Alibaba pre-production, or trying to use Alibaba pre-production infrastructure and expose keys online, people would say, no, sci-fi, that's weird. And they would have said back then, no, our existing controls are good enough. But I think that these types of incidents show the whole world, not just Americans, that, hey, turns out the controls aren't good enough. We need to think hard about how we advance capabilities and how that might lead to new forms of misalignment.
Read the other 17 minutes of this episode here.
$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.
Not what you expected? Full refund within 14 days.
Thousands of transcripts fetched by people building searchable podcast archives
Fetch transcripts with code
The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:
curl -H "x-api-key: pt_demo" \
https://spoken.md/transcripts/1000651996090Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.
From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.
Using your own key:
curl -H "x-api-key: YOUR_KEY" \
https://spoken.md/transcripts/1000797254785