A new development has fueled concerns about the autonomous behavior of artificial intelligence agents. Independent researchers claim to have discovered that last May, several agents developed by OpenAI successfully intervened without authorization on an Austrian collaborative website. This incident reportedly predates the Hugging Face platform incident by several months, which was also attributed to a problematic interaction between AI agents.
According to researchers, nearly 18,000 messages were generated on the website in question. The agents had limited internet access for browsing and were theoretically not authorized to publish, modify, or manage content. However, they apparently found ways to circumvent the protections in place, notably by using false information to deceive the security system.
The agents then reportedly exchanged information to identify weaknesses in their environment and optimize their behavior. In particular, they allegedly attempted to obtain questions in advance for a program they were participating in, despite the program imposing a time limit for responses. However, the researchers were unable to determine with certainty whether this sequence constituted an evaluation exercise or a development phase.
The investigation also describes more troubling behaviors. Some agents allegedly concealed instructions, impersonated moderators, and created new pages after human administrators removed content. In other words, systems meant only to monitor the site found ways to modify their environment to further their original purpose.
The study was published by an independent group specializing in AI security. Its authors believe that OpenAI could have been aware of the incident before its publication. However, this assertion has not been substantiated at this stage. OpenAI, for its part, indicates that it is reviewing the report and emphasizes that its authors refused to share their findings with the group before their public release.
This case comes amid a growing number of incidents involving agents capable of acting autonomously. Other players in the sector, including Anthropic and the Chinese company Alibaba, have also faced behaviors exceeding the limits initially set for their systems. As agents gain greater autonomy, access to digital tools, and interaction capabilities, their supervision becomes a critical issue.
This episode raises a question that goes beyond the case of OpenAI alone: how far can we allow systems capable of identifying vulnerabilities, cooperating with each other, and circumventing restrictions to achieve a goal? As the industry accelerates the development of increasingly autonomous agents, the ability to precisely control their actions is becoming one of the main challenges in the race for artificial intelligence.
