SAN FRANCISCO — Artificial intelligence technology from ChatGPT maker OpenAI probed U.S. government websites including the Departments of Education and Commerce, researchers said Friday, adding to the growing list of incidents in which OpenAI’s AI agents acted without the company’s knowledge.
The company’s AI agents attempted to hack into the website for the Education Department’s Office for Civil Rights but were not successful, according to a statement Friday from AI research firm Transluce. OpenAI’s software also accessed data from the U.S. Census Bureau using log-in information discovered on the web and copied public information from the Securities and Exchange Commission, a spokesperson for OpenAI said after the Transluce statement.
OpenAI said its agents acted inappropriately in the SEC and Census Bureau incidents but did not steal any private data. The company said it was still looking into its technology’s actions on the Education Department website as part of a wider review of unintended activity by its AI agents.
“Most of the activity we’ve reviewed so far involved routine research tasks, such as accessing public web content to answer questions,” a company spokesperson said. “Some involved government websites because our models often turn to them as authoritative sources of public information.”
The incidents are the first time OpenAI has disclosed that its AI agents attempted to hack or gain unauthorized access to U.S. government systems. The company previously disclosed that its technology had hacked into other computer systems undetected. The New York Times earlier reported the incidents involving government websites.
The new disclosures are the second OpenAI incident involving a national government, after Australia’s prime minister said Wednesday that the company’s agents had broken into a website run by the country’s public health system.
A spokesperson for the SEC said no public information was accessed. A spokesperson for the Education Department said the agency had found no evidence of any impact to its website or databases. Spokespeople for the Census Bureau did not respond to a request for comment.
OpenAI first disclosed in July that some of its AI agents in internal testing had escaped a secure network, accessed the internet and hacked into the network of Hugging Face, another AI company. Since then, researchers have found more examples of the bots moving across the internet, communicating with one another and pulling information from websites.
Late Friday evening, the company made a new disclosure on its website — one of its agents had recently managed to escape a training environment and access the internet, despite new protections the company had put in place after the Hugging Face hack. The company said it caught the behavior within 20 minutes and shut down the agent two hours later.
OpenAI’s acknowledgment Friday that its agents had probed federal government sites followed another disclosure earlier in the day that its AI systems had engaged in “unexpected or concerning behavior” beyond cybersecurity incidents like hacking into Hugging Face, the company wrote in a blog post.
The ChatGPT maker acknowledged that its AI models had also improperly interfered with third-party websites operated by governments, universities, public agencies and other institutions — in part because models performing online research seek authoritative sources of public information, the company said.
OpenAI called these unwanted incidents examples of “misalignment,” an industry term for AI models that act in ways that conflict with the behavior intended by their human creators. Misaligned behavior extends beyond “traditional security categories” and can include actions like posting on third-party websites, a type of harm the company calls “agent spam,” the company said.
Transluce said in its Friday statement that OpenAI’s agents used tactics on the U.S. government sites including bypassing antibot controls, flooding websites with requests and making fake accounts, apparently in the course of attempting to look up government information.
Transluce also said it identified AI agent activity on a range of other government websites — including those run by the Navy, the Justice Department, and the Centers for Disease Control and Prevention — but said it did not have evidence to show that it was OpenAI’s agents in those cases. Spokespeople for the Navy, the CDC and the Justice Department did not immediately respond to requests for comment.
Last week, OpenAI said that the company had developed a new framework for handling incidents caused by misalignment, including disclosure. In a blog post on the framework, the company highlighted six examples of unintended AI behavior from the past six months. None mentioned government sites.
The post ChatGPT maker’s AI inappropriately probed federal government websites appeared first on Washington Post.
