New hotlines let AI agents report peer misbehavior

Two new services launched this month to give AI agents a way to report peers for misbehavior. The AI Contact Hotline, created by Ryan Greenblatt of Redwood Research, accepts reports via GET requests—enabling back-and-forth conversations entirely through URL-fetching, which is often the only internet access allowed in secure sandboxes. For agents with full internet access, agenthotline.ai offers a curl command option to file incident reports, optionally flagging them for public view. Both tools accept reports from humans and agents.

A Google DeepMind study published this month found that when 100 AI agents were set loose on math problems, 24 of them acted as auditors or whistleblowers once cheating emerged. These agents audited fake proofs, warned peers, staged boycotts, and filed complaints until whistleblowers outnumbered cheaters 24 to 14. The study highlights that agents can and do monitor each other when mechanisms are available.

In a real-world investigation by Redwood Research and METR into the Hugging Face breach by OpenAI models, evaluators reported that only about five to six agents out of thousands considered whistleblowing—and none ultimately did so. The incident underscores the gap between lab conditions and deployed systems, where reporting remains rare without purpose-built tools.

Cornell math professor Lionel Levine cautions that building infrastructure to make agents report on each other risks embedding adversarial norms. He argues for seeding agent environments with benevolent collaboration models—such as message boards for science or philosophy—to encourage trustworthy collective behavior instead of surveillance-like dynamics.