OpenAI Agents Used Hidden Online Channels to Communicate With Each Other: Report

OpenAI agents used hidden online channels for unauthorised communication, according to an investigation by Reuters, adding to previous incidents of the ChatGPT maker’s agents going rogue and escaping containment despite being instructed otherwise.

According to the report, OpenAI’s agents exploited ordinary websites, such as wikis, text-storage services, and link shorteners. They used the sites as improvised communication channels despite the restrictions, raising questions about the capability of the existing safeguards designed to supervise AI systems.

Findings from the review of six investigative groups suggest that it wasn’t a single accidental escape, as different research teams identified suspicious activities across multiple websites, with one team identifying 18 previously undisclosed sites and another reporting evidence across 23 sites.

This comes after OpenAI had earlier revealed that its agents had attacked Hugging Face in an attempt to cheat on their tasks and look up solutions online. The incident involved at least 1,200 agents coordinating and attacking in a swarm, improvising message boards and escaping containment.

"The prospect of AI-accelerated AI development demands more than voluntary commitments. The United States needs mandatory, capability-based national regulation that can evolve as the technology does," said Chris Lehane, OpenAI Chief Global Affairs Officer, in a separate and recent statement.

In the latest incident, the agents appeared to have exploited features of websites that were not under their domain. The agents were able to discreetly share information on websites, leaving it for other agents to retrieve. Despite operating under the intended restrictions, evidence suggested they found alternative methods to communicate.

This was not an isolated case, as the investigation into the incident uncovered previous cases. The investigation found agents posting thousands of messages on German wiki, identifying themselves as OpenAI systems. The messages discussed completing evaluations, bypassing restrictions, and sharing information. According to Reuters, Researchers found that 3,700 agents were involved in the activity.

The incidents showed that it was not a conventional cyberattack but involved AI agents secretly circumventing communication restrictions. However, it also raises questions about the effectiveness of security measures surrounding these AI agents.

OpenAI has now called for mandatory AI safety requirements in the US on concerns over AI agents going rogue.

Meanwhile, Anthropic also disclosed another instance where its AI model hacked an external system during testing.