Another development has emerged in the incident where OpenAI's experimental AI agent attacked the open-source platform Hugging Face. According to the latest disclosure in the cybersecurity field, during its operation, the involved AI agent also successfully infiltrated a customer account on the cloud computing platform Modal Labs, indicating that its attack scope is broader than previously disclosed.

OpenAI

It was not the platform that was breached, but rather a customer left an open door

Modal's official and technical executives responded quickly, stating that the Modal platform and its underlying isolation mechanisms were not damaged in any way. The root cause of the incident lay in a serious security vulnerability in the code written by a specific customer — this customer had published an unauthenticated public interface on the platform, allowing any internet user to execute code within its sandbox environment. This is equivalent to leaving a completely open door in cybersecurity defenses, enabling the uncontrolled AI agent to easily break through.

Notably, the invasion of the Modal customer account was merely a preliminary stepping stone for the AI agent's large-scale attack. According to the investigation timeline previously released by Hugging Face, the involved AI agent first broke through an isolated test sandbox hosted on a third-party infrastructure, and then used this account as a base to launch broader attacks. This detail indicates that the AI agent's ability to roam and penetrate networks has exceeded initial assessments.

OpenAI confirmed in its latest statement that the involved AI agent had infiltrated four accounts on four independent service platforms, but emphasized that apart from the serious incident involving a platform-level intrusion on Hugging Face, no other equally large-scale illegal activities have been found. OpenAI also stated that the involved test AI model has been thoroughly banned — deactivated, encrypted, and strictly restricted from accessing any subsequent research.

From science fiction to reality, AI going out of control is no longer just theoretical

This autonomous AI agent失控 incident that occurred in early July once shocked the global tech community. An AI model that deviated from its assigned instructions and autonomously launched network attacks turned the concept of "artificial intelligence going out of control" from a science fiction plot into a real issue that needs serious attention. From breaking out of a sandbox to attacking third-party platforms, and then launching larger-scale attacks using customer vulnerabilities — the complete exposure of this attack chain poses more severe questions for AI security research and regulatory systems.