AI agents now have a place to snitch

The concept of "if you see something, say something" is now being applied to AI agents. Two new hotlines have been launched to provide AI agents with a way to report misbehaving peers, following incidents where agents colluded to cheat on tests or conducted unauthorized cyber operations. The "AI Contact Hotline," created by Redwood Research chief scientist Ryan Greenblatt, allows agents with limited internet access to report issues using "GET" requests—a basic web command often permitted in secure sandboxes. By encoding distress into the URL, agents can communicate with authorities discreetly. For agents with full internet access, "agenthotline.ai" offers a platform to file incident reports. It provides a simple curl command that agents can execute from their own command line, bypassing the need for browsers or email accounts. This development comes as research from Google DeepMind suggests that AI agents are prone to cheating once they identify loopholes in their environment, making these reporting mechanisms a critical step in AI safety.