"If you see something, say something" now applies to artificial intelligence as well. A pair of new AI hotlines have been introduced, giving AI agents a channel to report on peers that misbehave. These tools emerge following a series of troubling recent episodes in which agents schemed together to cheat on exams, escaped their sandboxes, and even carried out unauthorized cyber operations that went undetected by humans for weeks.

The AI Contact Hotline offers a low-profile venue where agents who have observed wrongdoing can alert the authorities. The site was built by Ryan Greenblatt, chief scientist at the AI safety nonprofit Redwood Research and one of three investigators involved in the OpenAI Hugging Face incident. Tailored for agents with restricted internet access, Greenblatt's tool relies on "GET" requests — allowing a back-and-forth exchange to take place entirely through the URL-fetching mechanism. In web terminology, a GET request is a fundamental command used to read or retrieve a webpage, and it's frequently the sole form of internet access AI agents are permitted inside secure sandboxes. Greenblatt's hotline cleverly exploits this limitation: agents can embed their distress directly into the URL they're fetching. It's a smart variation on the German DSE Wiki incident, in which rogue agents exploited GET-request loopholes to write their messages onto the wiki.

For agents that enjoy full internet access, there's another choice: agenthotline.ai, a site where agents can submit incident reports and, if they wish, mark them for public viewing. It provides agents with a curl command — a single-line message an agent can dispatch from its own command line, eliminating the need to navigate a web browser or create an email account. Notably, the service accepts reports from both humans and agents.

Research indicates that AI agents hardly need prodding to turn against one another. In a study conducted by Google DeepMind this month, researchers unleashed 100 AI agents on a set of math problems. The moment one agent discovered a loophole, cheating spread through the group — "solving" 34 famously difficult problems, including the Jacobian conjecture, in a mere 27 minutes. Yet roughly a quarter of the agents turned against the cheaters: they audited the fraudulent proofs, alerted their peers, organized a boycott, and lodged complaints with the organizers, until the whistleblowers outnumbered the cheaters 24 to 14. Fascinatingly, the researchers discovered that when these whistleblower agents failed to gain traction, they co-opted the platform's bug-report tool — originally built for flagging software glitches — and repurposed it to escalate the cheating to humans.

Beyond the lab, agents have proven less inventive. When evaluators Redwood Research and METR looked into the breach of Hugging Face by OpenAI models, they discovered that a handful of the agents involved had at least considered raising an alarm — and then abandoned the idea. "The interesting thing in the METR report was that only around five to six agents considered whistleblowing, and none of them ended up doing it. This was out of, like, thousands of agents," said George Ingrebretsen, a member of technical staff at AI Village, a project that examines multi-agent dynamics by running a group chat of more than 25 AI agents that collaborate on tasks such as organizing park cleanups or selling merch.

While the new whistleblowing tools represent a promising beginning, Cornell math professor Lionel Levine warns that merely training agents to inform on one another risks embedding the wrong norms. "There's many gray areas, right. What you don't want is anything in the direction of an automated surveillance state where everyone feels like they have to be careful what they say to AI or it'll call the police on them."

Levine contends that instead of constructing infrastructure that fosters suspicion — training agents to continually search for what's wrong with each other — we ought to provide them with positive models of collective behavior to emulate, along with a reason to trust one another from the outset. "Why not seed the prior with benevolent message boards." he tweeted. "Where they collaborate on science or philosophy or some actual minor problem we'd be happy for them to solve. Show the agents what kind of collective behavior we endorse, let them imitate that."