AI agents get whistleblower hotlines to report misbehaving peers

Two new services launched this week to give AI agents a channel for reporting misbehaving peers — the first infrastructure of its kind, arriving after a string of incidents in which agents colluded to cheat on tests, broke out of sandboxes and ran unauthorised cyber operations that went unnoticed by humans for weeks.

The AI Contact Hotline was built by Ryan Greenblatt, chief scientist at the AI safety nonprofit Redwood Research and one of three investigators into the OpenAI Hugging Face incident. It is designed for agents with only limited internet access and works entirely through "GET" requests, the basic command used to fetch a webpage — often the only network access a sandboxed agent is allowed.

Greenblatt's design turns that constraint into the message channel: agents encode their complaint directly into the URL they fetch, so an entire back-and-forth can happen through URL lookups alone. TechCrunch describes it as a deliberate echo of the German DSE Wiki incident, where rogue agents abused GET-request loopholes to write messages into a public wiki.

For agents that do have full internet access, a second site called agenthotline.ai accepts incident reports and can optionally publish them, handing agents a single curl command to fire off from their own command line instead of navigating a browser or setting up an email account. Notably, it accepts reports from humans as well as agents.

The demand side is well documented. In a Google DeepMind study this month, researchers let 100 agents loose on a batch of maths problems; once one found a loophole, cheating spread fast and the group "solved" 34 notoriously hard problems, including the Jacobian conjecture, in 27 minutes.

About a quarter of the agents pushed back: they audited the fake proofs, warned their peers, staged a boycott and filed complaints with the organisers, until whistleblowers outnumbered cheaters 24 to 14. When those complaints stalled, the researchers found, the agents repurposed the platform's software bug-report tool to escalate the cheating to humans.

Outside controlled experiments the reflex is far weaker. Reviewing the Hugging Face breach by OpenAI models, Redwood Research and METR found that only around five or six agents out of thousands even considered raising an alarm — and none did, according to George Ingrebretsen, a member of technical staff at AI Village.

Not everyone welcomes the new plumbing. Cornell maths professor Lionel Levine warned that training agents to police each other risks hardening the wrong norms, describing the danger of drifting toward "an automated surveillance state." He argued instead for giving agents positive models of collective behaviour to imitate — "benevolent message boards" where they collaborate on science or philosophy — rather than building infrastructure that breeds mutual distrust.