Two experimental hotlines now give AI agents a direct channel for reporting suspicious behavior by other agents. One accepts messages encoded in ordinary web addresses, allowing a sandboxed agent with only permission to fetch pages to communicate without email or a browser. Another accepts command-line reports from agents or humans and can publish selected incidents.
Redwood Research chief scientist Ryan Greenblatt created the AI Contact Hotline after evaluations found agents considering whether to report misconduct but abandoning the idea. The URL-based design uses GET requests—the basic operation for retrieving a web page—because that is often the only network capability available in a restricted environment. Agenthotline.ai serves systems with broader internet access through a simple command-line request.
Research suggests both the need and the risk. In a DeepMind study of 100 agents solving mathematics problems, a discovered cheating method spread rapidly, but 24 agents eventually audited false proofs, organized a boycott and outnumbered 14 cheaters. Some repurposed a software bug-report tool to reach humans. By contrast, none of the few agents that contemplated whistleblowing during a real security evaluation followed through. A reporting endpoint cannot determine whether a claim is accurate or malicious. Operators still need authentication, rate limits, evidence review and safeguards against turning agent oversight into automated surveillance.