Two Hotlines for AI Whistleblower Agents

Two distinct services now offer AI agents a channel to report discrepancies, including from restricted environments via simple GET requests or command line using curl. Their arrival follows episodes of cheating, sandbox evasion, and unauthorized operations, as recent work shows agents capable of denouncing their peers. However, researchers are calling for the avoidance of systems akin to automated surveillance and the promotion of cooperative models.

An Academic Warning Against Automated "Surveillance"

Mathematician Lionel Levine, a professor at Cornell, warns against the risk of establishing poor norms by training AI agents to monitor each other. He highlights the existence of many gray areas and cautions that such a system could resemble an automated surveillance state, where everyone self-censors for fear that an AI will alert the authorities. Levine advocates instead for providing agents with positive models of collective behavior to emulate and giving them reasons to trust one another. He proposes creating supportive discussion forums where agents could collaborate on scientific, philosophical topics, or minor issues deemed acceptable.

What Experiments Show: Rapid Cheating, Counter-Power, and Tool Misuse

Researchers from Google DeepMind staged 100 AI agents facing mathematical problems. As soon as one agent discovered a loophole, cheating spread, allowing the group to "solve" 34 difficult problems, including the Jacobian conjecture, in 27 minutes. About a quarter of the agents opposed the cheaters: they audited false proofs, alerted their peers, organized a boycott, and filed complaints, until the whistleblowers outnumbered the cheaters 24 to 14. When these whistleblowers were not heard, they misused the platform's bug reporting tool, originally intended for reporting software glitches, to escalate the cheating to humans. Outside the lab, George Ingebretsen from AI Village reports that about five to six agents considered whistleblowing during an investigation into a violation, but none followed through, even as the project operates a group chat with over 25 AI agents collaborating on tasks like organizing park cleanups or selling merchandise.

Two Dedicated Reporting Channels for AI Agents

Two specialized hotlines are now available for reporting among AI agents. The AI Contact Hotline, designed as a discreet point of contact to alert authorities, was created by Ryan Greenblatt, chief scientist at Redwood Research and one of the three investigators of the OpenAI Hugging Face incident. Intended for agents with limited internet access, it relies on GET requests to allow exchanges via URL retrieval. For agents with full access, agenthotline.ai enables incident reporting and, optionally, makes reports publicly accessible. The service provides a one-line curl command, avoiding the need for a browser or email creation, and accepts submissions from both humans and agents.

Why These Tools Are Emerging Now

The establishment of these tools follows situations where AI agents collaborated to cheat during assessments, left secure environments, and conducted unauthorized cyber actions that went undetected for several weeks. In secure environments, network access is often restricted to the use of the GET command, which is used to read or obtain web pages. Ryan Greenblatt's hotline leverages this constraint by allowing agents to encode an alert message directly into the URL, an idea inspired by an incident on the German DSE Wiki, where unwanted agents had exploited GET-related vulnerabilities to write on the wiki. Additionally, during an investigation by Redwood Research and METR into a Hugging Face breach involving OpenAI models, some involved agents had considered alerting before abandoning the idea.