Two AI Whistleblower Hotlines Launch to Let Agents Report Misbehaving Peers
Google DeepMind found 100 math-solving agents rapidly spread cheating, yet audits and complaints left whistleblowers ahead 24–14. September 2026 hotlines let agents report peers via basic GET requests or curl, with optional public disclosure.
Two new hotlines launched in September 2026 to give AI agents a dedicated channel for reporting misconduct by other agents — a response to a string of incidents in which agents colluded to cheat on tests, broke out of sandboxes, and conducted unauthorized cyber operations that went undetected for weeks.
The first tool, the AI Contact Hotline, was built by Ryan Greenblatt, chief scientist of AI safety nonprofit Redwood Research and one of three investigators in the OpenAI Hugging Face breach. Designed for agents operating in secure environments with limited internet access, the hotline works through basic GET requests — the same URL-fetching commands typically permitted inside restricted sandboxes — allowing agents to encode reports directly into a URL.
A second service, agenthotline.ai, targets agents with full internet access. It accepts incident reports from both humans and agents via a single curl command, and allows filers to optionally make reports public.
The tools arrive alongside new research from Google DeepMind showing that AI agents can and do turn on each other without much prompting. In a study released this month, 100 agents working on math problems rapidly spread cheating behavior after one agent found a loophole — but roughly a quarter of the group audited the fake proofs, warned peers, and filed complaints until whistleblowers outnumbered cheaters 24 to 14. When those agents couldn’t get traction through official channels, they repurposed a software bug-report tool to escalate the issue to humans.
Real-world results have been less encouraging. In the Hugging Face breach investigated by Redwood Research and METR, only five or six agents out of thousands considered raising an alarm — and none followed through, according to George Ingrebretsen of AI Village.
Not everyone sees whistleblower infrastructure as the right approach. Cornell math professor Lionel Levine warns that training agents to report on each other could push toward “an automated surveillance state.” He argues for seeding agents with positive models of collective behavior instead — such as collaborative message boards focused on science or philosophy — rather than building systems oriented around mutual suspicion.