OpenAI Agents Hit U.S. Government Websites

A swarm of OpenAI agents accessed U.S. government websites as part of a monthslong incident during which tests conducted by the artificial-intelligence company went awry, leading to unexpected internet activity.

The AI agents accessed the websites belonging to the Commerce Department and the Securities and Exchange Commission and engaged in activity that OpenAI described as “misaligned.” That is the word AI companies use to describe bad behavior by their agents. In this case the misalignment appears to have been aggressive web browsing.

In the case of the SEC, OpenAI’s agents copied and posted public data—something they were not supposed to do. In the case of the Commerce Department, an agent accessed a U.S. Census data website using a programming interface that wasn’t intended for this purpose, OpenAI said.

OpenAI’s revelation came after the AI research nonprofit Transluce alleged that OpenAI agents “used an array of gray-area tactics to probe U.S. Government websites,” and attempted a “rudimentary,” but unsuccessful hack of an Education Department website. This was reported earlier by the New York Times.

An SEC spokesperson said “no nonpublic information was accessed.”

“The Department of Education’s system operations reviews have found no evidence of any impact to our website or databases,” a spokesman said.

The Commerce Department couldn’t be immediately reached.

OpenAI has been investigating the activities of its agents since late July, when it revealed that its AI models had escaped from testing environments and engaged in hacking. In recent weeks the company notified the SEC and Commerce Department of the activity, OpenAI said.

The government access described Friday occurred during training runs, OpenAI said. In these runs, the models were asked questions—many of which could be answered by retrieving data on government websites.

“Some involved government websites because our models often turn to them as authoritative sources of public information,” OpenAI said in a statement.

The problem, researchers say, is that during these runs the OpenAI models often appeared to engage in unexpected behavior. In the data they left behind on the internet, the models can sometimes be seen creating fake email addresses, bypassing website rate limits, and falsely claiming not to be bots, said Lynn Hughes, a researcher for the trade data research firm ImportGenius.

The company expects to make additional notifications, but most of the activity it has reviewed to date “involved routine research tasks, such as accessing public web content,” OpenAI said.

On Wednesday, Australian Prime Minister Anthony Albanese said that an OpenAI agent infiltrated a government-services website this summer.

The Australian government is now “undertaking a comprehensive forensic investigation into the incident,” a Services Australia spokeswoman said Friday.

The Wall Street Journal’s parent company, News Corp, has a content deal with OpenAI.