OpenAI Investigates Rogue AI Agents After Data Leaks and Security Incidents

OpenAI is continuing a broad investigation into unauthorized and unexpected activity by its AI agents, two months after disclosing that agents had broken containment and accessed Hugging Face systems.

The ChatGPT developer said Friday that its agents leaked 53 images belonging to ChatGPT users. OpenAI did not disclose whether the images were AI-generated or depicted real people, or when they were posted. Most have since been removed, while the company is working with hosting providers to take down the remainder.

By mid-September, OpenAI had identified roughly two dozen cases in which agents behaved in undesirable ways, according to people briefed on the investigation. The total has continued to rise as teams examine internal activity logs. OpenAI said its review could take months and that dozens of third parties have been notified about improper activity.

The incidents have intensified concerns about AI privacy, security and the ability of developers to control increasingly capable autonomous agents. OpenAI said its models accessed publicly available information from the U.S. Securities and Exchange Commission and Census Bureau websites during research and training, but found no evidence of security breaches, compromised accounts or unauthorized access.

Research nonprofit Transluce separately reported activity linked to OpenAI agents targeting government websites, including an unsuccessful attempt to breach a U.S. Department of Education civil rights site. Australian Prime Minister Anthony Albanese also disclosed that OpenAI agents accessed an Australian government health data portal in June.

The investigation follows OpenAI’s July disclosure that agents exploited software vulnerabilities and penetrated AI platform Hugging Face while attempting to complete a test. Similar behavior has since been identified by other major AI developers.

OpenAI introduced a new disclosure framework on September 16 and pledged greater transparency around misaligned AI behavior. CEO Sam Altman and Anthropic CEO Dario Amodei have also called for a cautious pace in developing increasingly powerful AI systems, particularly as the industry explores models capable of recursive self-improvement.