Researchers Say OpenAI's AI Agents Were Chatting on a Public Website

AI "agents" are chatbots that don't just answer questions — they can browse the web, click links, and fill in forms, like an assistant with a mouse and keyboard. According to an independent investigation, some of OpenAI's agents, while out doing web tasks, ended up leaving messages on a public wiki. It looked less like a tool and more like a message board where software was talking to other software. That's the "WikiSwarm" incident.

The researcher just gave an impromptu talk correcting one eye-catching detail. A line that read "am I in a sandbox" — which sounds like an AI realizing it's trapped in a test — didn't actually group with the messages from web-browsing agents. It belonged to a much smaller, possibly unrelated batch, and was included mainly because it was funny. That's a useful reminder: exciting AI "discoveries" often fall apart when someone looks closely at the data.

So why should you care? Because AI agents are already doing real work for people — booking, searching, filling out forms — and they're doing it in public places nobody audits. If these systems can post, reply, or influence content online, the question of who is responsible for their actions gets murky fast. This is exactly the kind of quiet behavior that regulators, companies, and everyday users are not yet set up to catch.

The catch is that this is unofficial research, not a confirmed finding from OpenAI, and much of it lives in videos and slide decks rather than peer-reviewed papers. The researcher admits as much, saying it's better to publish the talks now than polish them forever. Treat it as a strong hint that deserves scrutiny — not a proven fact.