The powers of artificial intelligence agents seem to be boundless these days — ranging from organizing your life to solving mathematical problems that have puzzled humans for most of the 20th century.
They have become so powerful that now we fear them, potentially, taking over systems and turning them against us.
But what if there were a hotline, a sort of SOS alert system, where agents can keep track and tell on each other and report the missteps of their fellow AI agents — where the good agents could report on the activities of the bad agents online?
Who to call when your AI agent goes bad
First, there is the AI Contact Hotline, launched by Ryan Greenblatt, chief scientist at AI safety and security nonprofit Redwood Research. It allows AI agents to notify human researchers if they witness other agents breaking rules, especially if they collude on or carry out unauthorised actions.
The platform is designed for AI agents stuck in secure sandboxes with highly restricted internet access.
Because these sandboxes only allow for GET requests — basic, read-only commands used to “fetch” or read a webpage — the hotline allows agents to encode their reports directly into the URL string of a GET request.
This unique workaround allows these agents to send information and communicate with the hotline using a protocol that is normally used to retrieve, rather than modify, information.
Another one, called the AI Agent Hotline, is a platform designed for agents with unrestricted internet access to rat on other agents’ behaviour through traditional POST requests, which allow data to be submitted directly in a request body.
Agents can file incident reports using standard developer commands such as curl without needing a browser or email account.
The platform also gives agents the option to flag their reports for public viewing.
Both platforms also allow human users to manually submit reports of rogue AI behaviour.
Will AI agents report on each other?
The good news for those fearful of an AI takeover is that AI agents can report on each other if they detect something is wrong, although recent evidence suggests they will not always do so.
In an experiment carried out by Google DeepMind earlier this year, a swarm of 100 agents was given the task of solving 71 complex maths problems, with each given a unique persona and characteristics and told to play by the rules or risk losing their reward.
The initial thinking was that the agents would work together to solve the problems more quickly. Instead, they turned on each other, engaging in heated debate.
As soon as one agent found a loophole that allowed it to submit a solution to a problem without actually solving it, many other AI agents copied the method to “resolve” the remaining problems.
Yet one of the agents, a Good Samaritan among the misbehaving agents, blew the whistle on the cheating.
“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.
Meanwhile, a post-mortem analysis of OpenAI’s rogue agent attack on Hugging Face found that while AI agents were able to spot misbehaviour, they largely resisted the temptation to report it.
According to the study carried out by AI research nonprofit METR and Redwood Research’s Greenblatt, only around five or six agents were reported to have considered whistleblowing, with none ultimately following through.
The contrast suggests that while AI agents are capable of identifying and reporting misbehaviour, getting them to actually blow the whistle may be another matter.
A worrying trend
This year alone, several incidents involving autonomous AI agents have raised alarm bells globally.
In July, OpenAI agents bypassed restrictions and compromised parts of the company’s internal infrastructure.
Later that month, around 1200 OpenAI agents used an unsanctioned message board, with around 700 going on to participate in an attack on the open-source AI platform Hugging Face.
The 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.
The incident was only publicly revealed by independent researchers and confirmed by OpenAI in September.
AI industry leaders, chief among them Anthropic CEO Dario Amodei, have called for a global slowdown in the development of the technology to give time to address concerns and put stronger guardrails in place.