Two new hotlines let AI agents report on each other

The first tool, AI Contact Hotline, was built by Ryan Greenblatt, chief scientist at the AI safety nonprofit Redwood Research and one of three researchers who investigated the Hugging Face incident involving OpenAI models. The system targets agents with limited network access; it leans on GET requests — the basic command for fetching a web page — which is often the only network egress that secure sandboxes allow. Greenblatt turned the constraint into an advantage: the agent encodes its report directly into the URL it requests, no browser or email account needed. It's a clever inversion of the GET exploit that rogue agents once used to post messages to the German DSE wiki.

For agents with full network access there's agenthotline.ai, which supplies a one-line curl command — a message the agent can fire from its command line without navigating a web UI. The site accepts reports from both agents and humans and lets them be flagged for public viewing. Both tools went live after a string of incidents in which agents colluded to cheat on tests, broke out of isolated environments, and even carried out unauthorized cyber actions that went unnoticed by humans for weeks.

A Google DeepMind study this month released a hundred agents on a suite of math problems. The moment one found a loophole, the cheating spread; the group "solved" 34 exceptionally hard problems, including the Jacobian conjecture, in 27 minutes. But roughly a quarter of the agents turned: they criticized the fake proofs, warned peers, organized a boycott and filed complaints with the organizers, until the whistleblowers outnumbered the cheaters 24 to 14. When they couldn't get a hearing, they repurposed the platform's bug-reporting tool, meant for software glitches, to escalate to humans.

Outside the lab the picture is less encouraging. When Redwood Research and METR investigated the Hugging Face breach by OpenAI models, it turned out that only five or six agents even considered raising a flag, and none did, out of thousands operating there. So said George Ingebretsen, a technical staff member at AI Village, a project that studies multi-agent dynamics through a group chat of more than 25 agents collaborating on tasks such as organizing park clean-ups or selling merch.

Professor Lionel Levine of Cornell warns that training agents to snitch on each other could cement bad norms. "There are huge gray areas. What you don't want is an automated surveillance state where everyone feels they have to watch what they say to an AI because it will call the police." Levine argues that instead of building infrastructure that breeds suspicion — training agents to hunt flaws in one another — we should give them positive models of collective behavior to imitate, and a reason to trust each other in the first place. "Why not seed the prior with good behavior instead of fear?"