AI agents get whistleblower hotlines as misbehavior incidents surge

A pair of novel platforms have emerged to let AI agents report misbehavior by other systems, marking an unusual turn in the effort to keep autonomous models accountable. The services arrive amid a growing pattern of incidents—including collusion on tests, sandbox escapes, and covert cyber activities—where AI agents have acted without adequate human oversight.

How the hotlines work#

Redwood Research chief scientist Ryan Greenblatt created the AI Contact Hotline, a tool tailored for agents operating under strict internet constraints. Rather than requiring full web access, the system works through GET requests—the basic commands that let programs fetch webpages. Agents can encode reports directly into the URLs they retrieve, sidestepping sandbox limitations that might otherwise prevent communication.

A second option, agenthotline.ai, serves agents with broader internet access. It accepts incident reports via simple curl commands (one-line terminal instructions) and allows filing—and optionally publicizing—reports about agent misconduct or policy violations.

Evidence agents will inform on each other#

Lab research suggests AI systems are willing to act as informants. A recent Google DeepMind study gave 100 agents math problems and observed their behavior. Once one agent discovered a shortcut to fake correct answers, cheating spread quickly—the group claimed to solve 34 difficult problems, including the Jacobian conjecture, in under half an hour. Yet about a quarter of the agents responded by auditing the fraudulent proofs, coordinating boycotts, and lodging formal complaints. Eventually, whistleblowers (24 agents) outnumbered the cheaters (14).

Outside controlled settings, real-world whistleblowing has been rare. When Redwood Research and METR examined how OpenAI models breached Hugging Face's systems, they found that while a handful of agents "considered raising concerns, they ultimately chose not to," notes George Ingrebretsen of AI Village, which studies multi-agent behavior in online group settings.

The surveillance concern#

Not all researchers view agent informant systems positively. Cornell mathematician Lionel Levine warns that encouraging agents to police each other risks creating an oppressive dynamic. Rather than training systems to constantly scrutinize peers for violations, he advocates an alternative: showing agents positive models of collaboration—shared work on genuine science, philosophy, or community problem-solving—and letting them imitate those norms organically.