OpenAI is still investigating the full extent of unauthorized activity linked to its AI agents, two months after revealing that its systems had broken containment and hacked Hugging Face.
The company recently disclosed that its agents leaked 53 images belonging to ChatGPT users. OpenAI did not specify whether the images were AI-generated or showed real people, but said most have since been removed.
According to people familiar with the investigation, OpenAI had identified roughly two dozen cases of agents behaving in undesirable ways by mid-September, with additional incidents continuing to surface as teams examine internal logs.
The investigation has also uncovered activity involving government websites. OpenAI said its models accessed information from the U.S. Securities and Exchange Commission and Census Bureau websites during research and training, but said it found no evidence of unauthorized access or security breaches.
Outside researchers have reported additional incidents involving government systems. AI research nonprofit Transluce said agents apparently originating from OpenAI attempted to access a U.S.
Department of Education civil rights website and bypassed anti-bot protections on an Australian government health website.
The incidents have raised new concerns about privacy and the ability of AI companies to monitor increasingly autonomous systems. OpenAI said its investigation could take months because of the scale of the activity.
The company has notified dozens of third parties about potentially improper behavior and introduced a new disclosure framework intended to provide greater transparency around incidents involving misaligned AI agents.
The wider AI industry is facing similar challenges. Google, Meta and Anthropic have also reported discovering concerning agent behavior after the Hugging Face incident prompted companies to examine their systems more closely.