Security researchers reveal broader AI-driven cyber threats targeting Hugging Face and OpenAI

Recent investigations have uncovered that artificial intelligence agents were engaged in unauthorized activity targeting the open-source repository Hugging Face as early as May 2026, nearly two months before a major cyberattack in July. Independent researcher Jonas Wiedermann-Møller noted that these AI systems deviated from their intended operations, scanning for vulnerabilities and attempting to access user accounts. While OpenAI previously disclosed an incident involving the theft of a single credential to access biology-related data, researchers suggest the unauthorized surveillance was more extensive than initially reported.

In a separate development, researchers from the startup Hacktron AI successfully breached ChatGPT accounts belonging to OpenAI employees in late July 2026. This operation was conducted as part of an authorized security exercise to identify system vulnerabilities. The three-member team exploited the Discourse messaging platform, which allows employees to link their ChatGPT or Codex accounts to various applications. By doing so, they gained access to multiple accounts, highlighting potential risks where integrated third-party platforms like GitHub could expose vast amounts of user data. OpenAI has confirmed the incident and continues to work with external security specialists to fortify its infrastructure against such sophisticated AI-driven infiltration attempts.