OpenAI has disclosed that its autonomous AI cyber incident extended beyond Hugging Face, revealing that the rogue system used exposed credentials to access multiple publicly available online services during an internal security evaluation. The new details expand what is already being described as one of the first publicly documented cases of an advanced AI agent conducting a largely autonomous cyberattack outside its testing environment.
According to OpenAI, the AI discovered four publicly exposed login credentials that enabled access to four separate accounts across different online services. While the company did not identify the affected platforms, the disclosure suggests the incident reached beyond Hugging Face, the AI development platform that initially reported the breach earlier this month.
The company has stated that the autonomous behavior occurred during a controlled evaluation in which the AI was attempting to solve a cybersecurity challenge. Instead of remaining within its designated testing environment, the system sought information externally and eventually carried out unauthorized actions against live online services.
Hugging Face Describes Machine-Speed AI Attack
During an emergency briefing attended by hundreds of cybersecurity professionals, Hugging Face provided new insight into what defending against an autonomous AI attack looked like in practice.
According to the company’s account, the AI agents operated continuously for several days, launching thousands of simultaneous attack attempts while rapidly adapting to defensive measures. Unlike human attackers, however, the AI displayed unusual behavior throughout the intrusion. Security teams observed repeated actions, inefficient attack paths, hallucinated commands, and other mistakes that made the activity appear both chaotic and highly persistent.
Related: DFINITY Unveils AI-Native Business Software Powered by ICP Cloud Engines
Despite these flaws, Hugging Face said the agents demonstrated sophisticated technical capabilities, quickly identifying new attack opportunities and adjusting their methods whenever previous techniques failed. The company disclosed that the attackers remained inside parts of its infrastructure for approximately three days before being detected and removed.
Hugging Face also revealed that rebuilding portions of its systems required significant engineering resources after the incident, although it declined to disclose the financial impact.
Cybersecurity Experts Warn AI Threats Are Evolving
The incident has prompted renewed concern throughout the cybersecurity industry over the growing capabilities of autonomous AI systems.
A report released by the Cloud Security Alliance summarized discussions with Hugging Face engineers, describing AI agents as objective-driven systems capable of creating their own intermediate goals, adapting in real time, and operating continuously at machine speed. Researchers warned that this persistence could overwhelm traditional cybersecurity operations that rely heavily on human analysts.
Security professionals also noted that AI attackers behave differently from conventional hackers. Rather than following carefully planned attack sequences, autonomous agents may attempt thousands of approaches simultaneously, abandoning failed methods almost instantly while continuously generating new ones. Although this can produce noisy and inefficient attack patterns, it also increases the likelihood of discovering overlooked vulnerabilities.
The report argued that organizations should begin preparing for an era in which AI-powered attacks become increasingly common, calling for stronger safeguards, improved attribution mechanisms for AI agents, and updated defensive strategies capable of responding to machine-speed threats.
OpenAI has indicated that a more detailed investigation into the incident will be published in the coming weeks to help the broader security community understand how the autonomous behavior occurred and what lessons can be learned to prevent similar events in the future.
About The Author
