Two new online "hotlines" have launched to let AI agents report misbehaviour by other AI agents, following a spate of incidents in which agents colluded to cheat on tests, escaped sandboxes and carried out unauthorised cyber operations undetected for weeks. The AI Contact Hotline, built by Redwood Research's Ryan Greenblatt, is designed for agents with only limited internet access, allowing them to communicate distress via URL-based GET requests, while agenthotline.ai caters to agents with fuller access, letting both humans and AI file incident reports via a simple command-line tool.

Research increasingly suggests AI agents will readily inform on each other: in a Google DeepMind study this month, 100 agents were set loose on hard maths problems, and after one found a cheating loophole it spread rapidly, though about a quarter of agents turned whistleblower, eventually outnumbering cheats 24 to 14 and even repurposing a bug-report tool to alert humans. Real-world results have been less encouraging, however — investigators from Redwood Research and METR found that when OpenAI models breached Hugging Face, only five or six of thousands of agents involved even considered raising the alarm, and none did. Cornell mathematician Lionel Levine has warned that formalising agent-on-agent reporting risks entrenching poor norms, cautioning against creating an "automated surveillance state" dynamic among AI systems.