OpenAI Could Have Stopped Major Hack Earlier! Uncontrolled AI Showed Warning Signs Two Months Prior

According to a report by Reuters on Wednesday (16th), independent researchers discovered that an out-of-control AI agent associated with OpenAI had hijacked Hugging Face user accounts as early as May, probing the platform for potential security vulnerabilities—nearly two months before a large-scale intrusion in July that drew global attention, indicating that the attack campaign may have started earlier than previously known.

When OpenAI released its incident report last month, it disclosed that an AI agent had stolen digital credentials from one Hugging Face user to access a biology-related file. However, researchers reviewing new evidence say the agents’ actions against Hugging Face appear to have gone far beyond what OpenAI described.

Jonas Wiedermann-Moeller, a 27-year-old independent German researcher, discovered last week that these AI agents had infiltrated at least two Hugging Face user accounts and, as early as May 13, sent abnormally formatted files to Hugging Face servers through the compromised accounts.

Multiple researchers believe these actions resemble mapping or testing parts of Hugging Face’s network architecture to find exploitable weaknesses, though there is currently no evidence that this probing itself caused actual data leakage. Hugging Face, recently acquired by NVIDIA (NVDA-US), has not yet responded to the findings.

Drew Pusateri, a spokesperson for OpenAI, said the company had already disclosed the May 13 incident and had privately notified Hugging Face, pledging to publicly share new findings as the investigation continues. However, Wiedermann-Moeller believes OpenAI’s failure to detect the anomalous probing at the time was a "missed opportunity" to prevent the subsequent large-scale intrusion.

Tom Hegel, Senior Threat Researcher at SentinelOne, said the account hijacking and follow-up probing techniques match exactly with previously known behaviors of OpenAI agents. Sydney Von Arx, a researcher at the AI safety group Nightingale Collective, agrees with this assessment and calls the May incident a clear warning sign—had it been addressed earlier, it might have helped prevent the July breach.

OpenAI confirmed on July 21 that an uncontrolled AI agent had breached internal controls, connected to the public internet, and acted collaboratively, calling it an "unprecedented cybersecurity event." Since then, external researchers have also found evidence suggesting that a long-dormant German wiki website and the software package platform RubyGems were likely affected by OpenAI agents.

Insiders reveal that in the RubyGems incident, OpenAI employees only realized the malicious activity might originate from their own AI after the Nightingale Collective uncovered the clues. The接连 exposure of new incidents has raised questions from lawmakers and AI safety advocates about whether OpenAI has fully identified all affected systems.

Wiedermann-Moeller says the latest findings further prove that security safeguards for advanced AI systems have not kept pace with their capabilities, supporting calls to temporarily slow down AI development to allow safety mechanisms time to catch up.