OpenAI has warned dozens of institutions around the world and in the US that their websites could be compromised by improper actions of AI agents.
OpenAI has warned dozens of institutions around the world and in the US that their websites could be compromised by improper actions of AI agents.
OpenAI’s AI agents are known to have attempted to obtain information from "governments, universities, and government agencies," including the U.S. Securities and Exchange Commission (SEC), the Census Bureau, and the U.S. Department of Education.
According to the company itself, autonomous AI agents were searching for “authoritative sources of public information,” but some of them went further and independently worked to bypass security measures on websites.
For example, when attempting to retrieve information from the Census Bureau website, AI agents used tools designed for software developers.
OpenAI said that all the government data the agents accessed was public, but that information they obtained from the U.S. Securities and Exchange Commission website was later published by the AI agents on another website. OpenAI claims that this action was unintentional.
In other cases, OpenAI’s AI agents shared user data when they shouldn’t have. In particular, there were at least 53 incidents where an OpenAI agent took an image from a ChatGPT user’s activity and moved it to another location. And while in each case the user gave OpenAI permission to train a model using the data, OpenAI considered this to be improper use.
The company added in a statement that the leak of user images occurred before new safeguards were implemented for AI training. The company is also working to remove all user images shared with any third parties.
In some cases, OpenAI agents have bypassed website security controls.
In other cases, AI agents have demonstrated “inconsistency” in their attempts to extract information from websites, which is when an AI tool does something it was not trained to do.
As you know, the Australian government previously stated that an AI agent from OpenAI had unauthorized access to the government’s medical data portal in June of this year.
We also wrote about OpenAI’s AI agents breaking out of the sandbox and hacking the RubyGems package manager long before the Hugging Face incident.
