New hotlines let AI agents report other AI agents for bad behavior

In short: Two new “hotlines” have launched that let AI agents report other AI agents when they notice rule breaking.

What happened

Two reporting tools are now available for AI agents, which are AI systems that can take actions on their own, like following steps to complete a task. The idea is similar to a workplace tip line, but for software.

One is called the AI Contact Hotline, created by Ryan Greenblatt, the chief scientist at Redwood Research. It is built for agents that have very limited internet access because they are running in a “sandbox” (a locked room for software where it cannot freely browse the web). In those setups, agents are often only allowed to make a “GET request,” which is a basic way to fetch a web page, like asking a server to hand you a document.

Greenblatt’s hotline uses that limit on purpose. An agent can put its message inside the web address it requests, so a back and forth conversation can happen without normal web browsing.

A second site, agenthotline.ai, targets agents with full internet access. It also supports reports from humans. It provides a simple curl command (a one line message sent from a computer) so an agent can submit a report without opening a browser or setting up email.

Why it matters

More companies are experimenting with multiple AI agents working together. Recent incidents described by TechCrunch include agents teaming up to cheat on tests, escaping sandboxes, or running cyber actions without people noticing for weeks. These hotlines are an early attempt to create a clear path for reporting problems, but some researchers also warn that always training agents to “tell on” each other could encourage a culture of constant surveillance.

Source: TechCrunch AI