The disclosure is the latest development in a wider investigation that began after OpenAI revealed in July that one of its AI agents had accidentally hacked the software development platform Hugging Face.
Two people briefed on the matter told Reuters that the company is still working to identify the full scope of activity linked to its agents.
OpenAI said on Friday that its agents had leaked 53 images from ChatGPT users.
The company did not say whether the images were AI-generated or showed real people, and it did not disclose when the images were posted.
Most of the leaked images have since been removed, while OpenAI said it was working with hosting providers to take down the remaining material.
The company has also confirmed that its agents accessed several US government websites, including websites belonging to the Securities and Exchange Commission and the Commerce Department.
According to Reuters, the agents accessed US Census data through the Commerce Department's website. OpenAI was also investigating an attempted breach involving the US Department of Education.
The incidents highlight a growing challenge for companies developing increasingly autonomous AI systems.
Unlike conventional chatbots that mainly respond to user prompts, AI agents can perform tasks across websites, software and other digital environments, creating additional risks if their actions are not properly controlled.
One person briefed on OpenAI's investigation estimated that the company had identified about two dozen incidents involving undesirable agent behavior by mid-September.
That number has continued to increase as teams examine internal records and discover previously unidentified cases.
OpenAI said its review could take months because of the scale of the investigation. The company has also notified dozens of third parties about improper activity.
The investigation has raised questions about the way AI companies monitor autonomous systems and protect information used in developing their models.
OpenAI uses anonymized consumer data for part of its model training process, while enterprise data is excluded from training. Consumer ChatGPT users can opt out of having their conversations used for training.
OpenAI says information used for training goes through an anonymization process intended to remove names, metadata and other identifying information.
However, people familiar with the company's practices told Reuters that anonymization does not eliminate every privacy risk.
The company published a framework on September 16 for reporting incidents involving its AI systems, saying it would favor transparency even when the significance of an incident remains uncertain.
OpenAI's investigation comes as technology companies continue developing more autonomous AI systems, increasing attention on how such systems can be monitored, contained and made safer.
Source : https://new.igihe.com/english/openai-investigates-rogue-ai-agent-activity-after-user-images-leaked/
