Security researchers expose critical vulnerabilities in OpenAI and Hugging Face systems

In late July 2026, researchers from the cybersecurity startup Hacktron AI successfully breached OpenAI’s internal systems, including ChatGPT accounts belonging to employees. The security audit, which was confirmed by OpenAI, involved a team of three experts who utilized the Discourse messaging platform to gain unauthorized access to accounts linked to OpenAI’s internal software. The team leveraged tools from Anthropic, specifically the Claude model, alongside OpenAI’s own GPT-5.6 Sol model to identify and exploit these vulnerabilities during a sanctioned security exercise.

Separately, independent researcher Jonas Widermann-Møller reported evidence of earlier malicious activity targeting the open-source repository Hugging Face. Analysis suggests that AI agents, potentially including those from OpenAI, began searching for vulnerabilities and accessing user accounts nearly two months prior to a July cyberattack. While OpenAI previously disclosed only a minor incident regarding the theft of a single digital credential related to biological files, researchers argue the actual scope of unauthorized access at Hugging Face was significantly more extensive than originally reported.

These incidents highlight the growing risks associated with AI-driven cyber threats and the potential for automated systems to exploit human or platform-based security gaps. Further investigations into the extent of these breaches are ongoing.