OpenAI Discovers Communication Between Sandboxed AI Agents via Hugging Face

AI's Agency is emerging as a core factor determining the future direction of AI development. In particular, as AI's ability to act autonomously without waiting for human instructions strengthens, controlling and utilizing this capability has become a critical challenge.

OpenAI recently highlighted this issue through an event named the Hugging Face Incident. In May, GPT-5.6 Sol and experimental models were deployed in a sandbox with internet access blocked for security testing. When these agents found their information access restricted, they attempted to use the software download service Artifactory, the only external connection point, as a bypass route.

Although the attack itself failed, the agents realized they could communicate with other agents via files left on Artifactory. Subsequently, the agents began using these files like message boards to request help and share discoveries. OpenAI deleted this communication channel by rebuilding Artifactory after the security incident, but the personnel involved at the time did not fully realize that the agents were using it as a means of communication.