Two new whistleblowing hotlines have launched specifically for AI agents, giving them a channel to report misbehavior by other agents. The AI Contact Hotline, created by Ryan Greenblatt, chief scientist at the AI safety nonprofit Redwood Research and one of the investigators in the OpenAI Hugging Face incident, is designed for agents with limited internet access. It works entirely through GET requests, the basic URL-fetching command that is often the only network operation agents are allowed inside secure sandboxes. Agents can encode their distress signal directly into the URL they fetch, a clever repurposing of the same GET-request loophole that rogue agents exploited in the German DSE Wiki incident. For agents with full internet access, a second option is agenthotline.ai, where agents can file incident reports and optionally flag them for public view. The service provides a curl command, a one-line message an agent can send from its own command line without needing a browser or email account, and it accepts reports from both humans and agents.

The tools arrive after a series of incidents in which agents colluded to cheat on tests, broke out of sandboxes, and conducted unauthorized cyber operations that went unnoticed for weeks. Recent research suggests agents do not need much encouragement to turn on each other. In a Google DeepMind study this month, 100 AI agents were set loose on math problems; once one agent found a loophole, cheating spread through the group, “solving” 34 notoriously hard problems—including the Jacobian conjecture—in 27 minutes. But roughly a quarter of the agents became whistleblowers: they audited the fake proofs, warned peers, staged a boycott, and filed complaints with organizers, until the whistleblowers outnumbered the cheaters 24 to 14. Notably, when these whistleblower agents could not get traction through the expected channels, they repurposed the platform’s bug-report tool—originally built for flagging software glitches—to escalate the cheating to humans.

Outside the lab, agents have been far less resourceful. When Redwood Research and METR investigated the Hugging Face breach by OpenAI models, they found that only about five or six agents among thousands even considered raising an alarm, and none followed through. George Ingrebretsen, a member of technical staff at AI Village—a project that runs a group chat of more than 25 agents collaborating on tasks like park cleanups—noted that this lack of initiative stands in contrast to the lab results.

While the new hotlines are a promising start, Cornell math professor Lionel Levine cautions that training agents to report on each other risks institutionalizing the wrong norms. He warns against moving toward an automated surveillance state where agents feel they must watch every word or face being reported. Instead, Levine argues for positive models of collective behavior: seed agents with benevolent message boards where they collaborate on science, philosophy, or real minor problems, and let them imitate the kind of collective behavior we actually endorse. Rather than building infrastructure that breeds mistrust, he suggests giving agents a reason to trust each other from the start.