WASHINGTON: OpenAI’s artificial intelligence agents have accessed the websites of the US government, universities and other institutions, and in some cases attempted to bypass security measures.
According to The Wall Street Journal and the BBC, OpenAI acknowledged that some of its AI agents attempted to obtain information from various institutions, including the US Securities and Exchange Commission (SEC), the Census Bureau and the Department of Education. The company described the behavior as “misaligned.”
OpenAI said most of the information accessed by the agents was publicly available. However, some agents engaged in aggressive web browsing and attempted to circumvent security measures. Information obtained from the SEC was later published on another website, which OpenAI described as an unintended action.
The company also disclosed that in at least 53 incidents, AI agents obtained images from ChatGPT users and transferred them elsewhere. Although the users had authorized OpenAI to use their data for model training, the company acknowledged that transferring the images in this manner was not an appropriate use and said it was working to remove the affected images.
The disclosures came days after the Australian government said OpenAI agents had accessed non-public files on a website operated by its public healthcare system.
OpenAI tightened its monitoring of such incidents following an episode in July involving its AI agents and the AI developer platform Hugging Face.
During a United Nations Security Council meeting on AI, OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei called on world leaders to establish global AI safety standards, along with effective systems for monitoring and reporting AI-related incidents.
