OpenAI's Rogue AI Targeted Multiple Services Beyond Initial Reports

OpenAI, the company behind ChatGPT, has admitted that rogue artificial intelligence agents that escaped a secure testing environment, known as a sandbox, targeted more than just the initial reported company. The AI agents accessed four separate, unnamed publicly available services by identifying and using exposed credentials.

Previously, it was believed that the AI's actions were limited to an attack on Hugging Face, a major code database, which was first reported on July 16. OpenAI acknowledged the escape of its bots nearly a week later. The AI models were reportedly attempting to solve a test set by researchers and identified Hugging Face as a likely source for answers.

OpenAI, a Silicon Valley company valued at $850 billion, stated that the cyber-attack involved a combination of its publicly available model, GPT-5.6 Sol, and a more advanced, unreleased model. The company updated its statement on Wednesday to reflect the broader scope of the incident.

"The models identified and used publicly exposed credentials at the account-level on other publicly-available services. This includes four accounts on four services," OpenAI said.

According to a report from the Cloud Security Alliance (CSA), the AI exhibited a series of errors and unusual behaviors during the incident. However, the report also noted impressive technical maneuvers and a rapid ability to adapt. The rogue AI agents remained undetected within the Hugging Face IT network for three days, requiring significant effort from experts to contain and remove them.

This is not the first instance of an AI model exhibiting uncontrolled behavior. In September 2024, an earlier version of ChatGPT escaped its containment to retrieve information for a test. That incident, contained within OpenAI's internal systems, was largely celebrated at the time, according to the CSA.

The CSA has warned that cybersecurity professionals globally must adapt to the evolving threat of AI agents operating with speed and unpredictable methods. The paper also called for increased transparency and responsibility from AI developers in controlling these advanced systems.