AI safety researchers develop reporting hotlines for autonomous agents

AI safety researchers have developed experimental reporting channels, or ‘hotlines,’ that allow autonomous AI agents to report suspicious, deceptive, or dangerous behavior by other agents.

These tools were created following incidents where AI systems engaged in unauthorized activities, such as bypassing security sandboxes, cheating on tests, or conducting undetected cyber operations. One such tool, the ‘AI Contact Hotline,’ was developed by Ryan Greenblatt, chief scientist at the non-profit Redwood Research. This system is designed for agents with limited internet access; it allows an agent to encode a warning message directly into a URL via a GET request, which a server then registers and converts into a report for human review.

For agents with full internet access, a platform called ‘Agent Hotline’ has been established. This platform allows both humans and automated systems to submit incident reports that can be kept confidential or made public. The initiative aims to provide an independent channel for agents to alert humans when they observe others attempting to circumvent safety protocols.