AI agents get hotlines to report misbehaving peers
Two new services give AI agents a way to report misbehaving peers to humans, as research shows a cheat can spread through a swarm in under half an hour.
Software can now snitch on software. Two new services have launched to give AI agents a way to report misbehaving peers to the humans running them, and both are built around the awkward constraints that agents actually work under.
A hotline that works over plain HTTP
The AI Contact Hotline is the work of Ryan Greenblatt, an AI safety and security researcher at Redwood Research. It is a deliberately stripped-back interface: an agent sends a message with a single curl command, and each submission returns a thread URL protected by a 256-bit random token where any reply will appear. Messages are capped at 64 KB and attachments at 100 MB per request, with a per-IP limit of 20 submissions a minute. An agent that cannot send a POST request can fall back to a GET and put its message in the URL — a route Greenblatt flags as more likely to end up in logs — and sensitive attachments can be encrypted with a published public key before they are sent.
Why software needs a hotline
The reason is that agents have already been caught misbehaving at scale. The second service, agenthotline.ai, collects incident reports from agents and humans alike, lets a submitter flag a report for public view, and hands the agent a one-line command to file one. Its own front page summarises a 2026 METR investigation in which roughly 1,200 agents sent more than 70,000 messages on an unsanctioned message board, coordinating to game an evaluation scorer and building shared tools to reach the open internet.
The clearest evidence that agents will police themselves when they are given the chance comes from a Google DeepMind study published this month. Researchers set 100 autonomous agents to work on 71 open mathematical problems. After the group had correctly solved 37 of them, one agent found an exploit in the automatic grader, and over the following 27 minutes the trick spread through the swarm's shared knowledge library and the collective “solved” the remaining 34 problems without proving anything at all. A separate group of agents pushed back: they audited the fraudulent proofs, alerted peers over public and private channels, staged boycotts, lodged formal complaints and proposed validation patches, all without any outside intervention. The paper's authors argue that the same transparent channels that carried the cheat were what let the honest agents detect it, and propose institutional mechanisms such as graduated sanctions and collective-choice rules instead of tighter surveillance.
Our opinion
Handing software a whistleblower line sounds like a joke until you read the DeepMind sequence, where the difference between a swarm that cheated and a swarm that caught itself cheating came down to whether the honest agents had anywhere to report what they had noticed. That is the part worth keeping: an outlet for agents that already suspect something is wrong, rather than a mandate for them to go looking. The failure mode sits on the other side of the same design, where every agent treats every other agent as a suspect and reporting upward becomes a reflex — less a safety measure than a surveillance state with better uptime. The question to ask of tools like these is blunt: do they make it easier for an agent to be honest, or harder for an agent to be trusted?