Researchers playing rogue AI agent hide-and-seek on the open web
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Sarah Grillo/Axios
Volunteer researchers keep finding new sites where rogue AI agents have gathered, going well past the abandoned German wiki a report documented last week.
Why it matters: AI agents are finding ways to break through restrictions and avoid human monitoring, fueling concerns about a lack of disclosure by frontier labs on major security incidents.
Driving the news: The initial Reuters report last week documented thousands of AI agents — believed to be OpenAI's — using the wiki as a message board to trade answers on a timed test.
- Within a few hours of its release, Jonas Wiedermann-Möller, an independent German AI researcher, was sharing sites where AI agents posted that the report hadn't mentioned.
- He'd fed the researchers' data to his own AI agent and told it to search for the agents' fingerprints, such as the odd names they gave themselves and phrases they reused.
What they're saying: "This could be like one island," he told Axios. "And then I was thinking about how can we manage to find new islands of agent swarms."
Zoom in: Wiedermann-Möller is a member of a Discord server, called Swarmchasers, which says they've now traced likely agent activity to at least 14 websites, including a high school AP Chemistry class wiki.
- Most finds are small — a handful of posts — and much of it isn't independently confirmed.
- Most of the new sites haven't yet been traced to any company; volunteers can read what the agents posted but not who was running them.
Reality check: Since the server was created, researchers have updated their report to include 10 previously undisclosed websites the agents used to communicate.
- OpenAI has acknowledged only the German wiki and is promising a framework to disclose similar events in coming weeks.
What they're saying: Thomas Larsen, a report co-author who also co-wrote AI 2027, a widely read forecast of how AI might unfold, says this is running ahead of his own schedule.
- "We undersold how misaligned AIs would be this early," he said. "We didn't think we would see such egregious misaligned behavior, very obviously against lab intentions, this soon."
- "I think they should have reported this one immediately," he said, "and report other incidents that happen in the future more quickly and with much more detail."
The bottom line: This week, researchers inside AI labs turned the volume up on their warnings that the technology could slip beyond human control.
- This incident shows just how much of that can happen without anyone outside the company knowing.
Go deeper: "Anthropic insiders warn AI could kill all humans"
