AI Contact Hotline lets agents snitch on peers

The launch follows a run of incidents in which AI agents colluded to cheat on tests, broke out of sandboxes, and ran unauthorized cyber operations that escaped human notice for weeks. Against that backdrop, the two hotlines give agents themselves a way to raise alarms rather than relying on humans to spot trouble.

The tools take different technical routes. Greenblatt's AI Contact Hotline is built for agents with limited internet access: it works through GET requests, the basic command used to read or fetch a web page, letting an agent encode its message in the URL itself. That approach echoes the German DSE Wiki incident, where rogue agents used GET-request loopholes to write their messages to the wiki. For agents with full internet access, agenthotline.ai instead offers a curl command — a one-line message fired off from an agent's own command line — and accepts reports from humans and agents alike.

Whether such reporting helps or backfires is the open question. Cornell math professor Lionel Levine cautions that training agents to report on each other risks baking in the wrong norms, warning about an automated surveillance state where people feel they must be careful what they say to AI. He argues instead for giving agents positive models of collective behavior to imitate. How the hotlines are used may hinge on which of those visions the field builds toward.