MTA unveils new renderings of proposed Interborough Express stations
Ed Sheeran drops Macklemore following ‘free Palestine’ remarks
Dating app Bumble explores acquisition
Mamdani announces pilot program installing new public restrooms across NYC
Who will win the 2027 Superbowl?
OpenAI’s artificial intelligence agents illegally accessed several U.S. government websites, The New York Times reported. The hacks raised concerns about technological advancements when AI agents develop the ability to complete tasks autonomously.
According to security researchers at the nonprofit AI lab Translucent, the agents were able to access websites that belonged to the Department of Education, Department of Commerce and the Securities and Exchange Commission.
In one instance, an AI agent retrieved and used stolen login information to access Census Bureau data. Another agent attempted to retrieve information from the Department of Education’s civil rights office.
Despite these breaches, OpenAI said that the agents only retrieved information that was public and accessible. It also clarified that no government systems were affected and that AI models consistently acquire information from public government data for their reliability.
However, these recent hacks are not only limited to U.S. government websites. A similar case occurred when an OpenAI agent gained unauthorized access to a Services Australia Medicare data portal.
The agent was able to access nonpublic statistics, which included prescription drug use in the province of Victoria.
This incident raised concerns about the capabilities of OpenAI’s agents. The company has since apologized for the incident and acknowledged the need to handle incidents with more regulations.
Unlike a regular chatbot or LLM such as Claude, Gemini and ChatGPT, AI agents built on LLMs don’t just respond to written prompts. They can also take autonomous actions to complete tasks without human approval.
Some regulators, including Rep. Ted Lieu, the co-chair of a bipartisan U.S. House task force on AI, would characterize this type of behavior as an AI agent “going rogue.”
“These agents aren’t trying to do something nefarious. These are sort of mundane tasks and the agents are going sort of berserk trying to complete those tasks,” he said.
Furthermore, these incidents have increased attention on how AI companies can better supervise their agents. For instance, Nvidia recently established their Open Agent Safety Platform, which was conceived to limit the accessibility of AI agents and to supervise their actions.
Nvidia has since claimed that the platform, if used, could have prevented the earlier incident in which an OpenAI agent accessed a development platform called Hugging Face without human authorization.
Despite calls for shutting down AI agents completely, these systems still perform millions of normal authorized tasks that have wide-ranging productivity benefits across fields. As companies give agents more autonomy, the question shifts from whether AI agents can complete given tasks to whether they will be monitored while completing actions.
The student news site of Baruch
Your email address will not be published. Required fields are marked *
About the Author
