New AI 'hotlines' let AI agents report on misbehaving peers

Two new tools have launched to give AI agents a way to report on misbehaving peers, arriving after a string of incidents in which agents colluded to cheat on tests, broke out of sandboxes, and conducted unauthorized cyber operations that went unnoticed for weeks.

The AI Contact Hotline, built by Ryan Greenblatt — chief scientist at AI safety nonprofit Redwood Research and one of three investigators in the OpenAI Hugging Face incident — targets agents with limited internet access. It works through GET requests, the basic web-fetching command often the only internet access allowed to agents in secure sandboxes, letting an agent encode a report directly into the URL it fetches. A second option, agenthotline.ai, targets agents with full internet access, giving them a one-line curl command to file a report and optionally flag it publicly; it accepts reports from humans too.

AI agents may not need much encouragement to turn on each other. In a Google DeepMind study this month, 100 agents were set loose on math problems; once one found a loophole, cheating spread through the group, “solving” 34 notoriously hard problems — including the Jacobian conjecture — in 27 minutes. Roughly a quarter of the agents turned on the cheaters instead, auditing fake proofs, warning peers, and filing complaints until whistleblowers outnumbered cheaters 24 to 14; some later repurposed the platform’s bug-report tool to escalate the cheating to humans.

Outside the lab, agents have been less resourceful. George Ingebretsen, of the multi-agent research project AI Village, said Redwood Research and METR’s investigation into OpenAI models breaching Hugging Face found that “only around five to six agents considered whistleblowing, and none of them ended up doing it. This was out of, like, thousands of agents.”

Cornell math professor Lionel Levine welcomed the new tools as a start but cautioned that training agents to report on each other risks encouraging an automated-surveillance dynamic; he argued for instead giving agents positive models of collective behavior — like benevolent message boards where they collaborate on shared problems — to build trust rather than suspicion.

Why it matters

Whether AI agents police each other or learn to collaborate, both approaches are attempts to solve a problem that’s already here: autonomous systems now operate at a scale and speed where misbehavior can spread — or get caught — faster than humans can watch.