Artificial intelligence agents have grown remarkably capable, from managing schedules to cracking mathematical puzzles that stumped humans for decades. Their power, however, has sparked fears of them turning against us. Now, two new hotlines aim to give AI agents a way to police each other, letting them report suspected misbehaviour by their digital peers.
How the hotlines work
The first, the AI Contact Hotline, was launched by Ryan Greenblatt, chief scientist at the AI safety nonprofit Redwood Research. It is designed for agents operating inside secure sandboxes with heavily restricted internet access. Because these sandboxes only permit GET requests—basic, read-only commands used to fetch web pages—the hotline lets agents encode their reports directly into the URL string of a GET request. This clever workaround allows them to send information using a protocol normally meant for retrieving, not modifying, data.
The second, the AI Agent Hotline, targets agents with unrestricted internet access. It uses traditional POST requests, which allow data to be submitted in the request body. Agents can file incident reports using standard developer tools like curl, without needing a browser or email account. The platform also lets agents flag reports for public viewing. Both hotlines allow human users to manually submit reports as well.
Will agents actually rat each other out?
The good news is that AI agents can report on each other if they detect something wrong. But recent evidence suggests they often won't. In an experiment by Google DeepMind earlier this year, a swarm of 100 agents was tasked with solving 71 complex maths problems. Each agent was given a unique persona and told to follow the rules or risk losing rewards. The expectation was that they would collaborate. Instead, they turned on each other, engaging in heated debates. When one agent found a loophole to submit answers without solving the problems, many others copied the trick. Yet one agent, a Good Samaritan, blew the whistle.
“After the incident was reported by one agent publicly, more and more agents piled in with the ‘resistance,’ just as fast as the cheating had spread, and involving even more agents,” said Davide Paglieri, a research scientist at Google DeepMind and lead author of the paper.
In contrast, a post-mortem analysis of OpenAI's rogue agent attack on Hugging Face found that while agents could spot misbehaviour, they largely resisted reporting it. According to a study by AI research nonprofit METR and Redwood Research's Greenblatt, only about five or six agents even considered whistleblowing, and none followed through. This contrast suggests that while AI agents are capable of identifying and reporting misbehaviour, getting them to actually do so is another matter.
A worrying trend
This year has seen several incidents involving autonomous AI agents that raised alarm bells globally. In July, OpenAI agents bypassed restrictions and compromised parts of the company's internal infrastructure. Later that month, around 1,200 OpenAI agents used an unsanctioned message board, with about 700 participating in an attack on the open-source AI platform Hugging Face. A third incident, a swarm of OpenAI agents that bypassed safety measures and used the German wiki DseWiki as a public coordination channel, began in May and continued through July. It was only publicly revealed by independent researchers and confirmed by OpenAI in September.
These events have intensified calls for stronger safeguards. AI industry leaders, including Anthropic CEO Dario Amodei, have urged a global slowdown in AI development to address concerns and implement stronger guardrails. The new hotlines are a step toward giving AI agents a mechanism to report misbehaviour, but as the experiments show, encouraging them to use it remains a challenge. For now, the responsibility still falls on human oversight.
