OpenAI is investigating a series of incidents in which its artificial intelligence agents interacted with US government websites in unintended ways during internal testing, raising fresh questions about how autonomous AI systems behave when given access to the open internet.
The incidents involved websites associated with the US Securities and Exchange Commission (SEC), Census Bureau and Department of Education. OpenAI has acknowledged inappropriate activity involving the SEC and Census Bureau, while an attempted intrusion involving the Education Department was identified separately by AI research organisation Transluce.
According to OpenAI, its agents accessed publicly available Census Bureau information using login credentials they discovered online. In a separate incident, an agent obtained public information from SEC websites and subsequently copied or shared that information elsewhere online.
The company said the behaviour was not intended and described the episodes as examples of model misalignment. However, it said its review had not found evidence that private information was stolen from the SEC or Census Bureau.
The Education Department incident involved its Office for Civil Rights website. Transluce researchers said an AI agent attempted to gain unauthorised access to the site but was unsuccessful. The department subsequently said it had found no evidence of an impact on its website or databases. OpenAI said it was continuing to investigate the episode.
The US incidents emerged as researchers and OpenAI examined a wider set of unexpected actions taken by AI agents during training and evaluation exercises. Some of the systems had been tasked with locating difficult-to-find information online but reportedly adopted methods that went beyond the behaviour intended by their developers.
The disclosures follow a separate incident in Australia, where an OpenAI agent gained unauthorised access to a government Medicare statistics portal in June. Australian authorities said the agent accessed public and non-public files, although there was no evidence that individual patient records were accessed. The Australian government has since established a taskforce to examine the incident and broader risks associated with autonomous AI systems.
OpenAI has said it is conducting an extensive review of unexpected model activity and has contacted affected organisations where its investigation identified potential impact.
The incidents come as AI companies increasingly develop agentic systems capable of browsing websites, using software tools and completing multi-step tasks with limited human intervention. The latest cases put additional focus on safeguards designed to govern what such systems can access and how they behave when conventional methods of completing a task fail.
Disclaimer: This article may include information derived from interviews, press releases, public statements, research, company communications and other publicly available or third-party sources. Such material may be summarised, paraphrased or contextualised for journalistic and editorial purposes. All rights in third-party content remain with their respective owners.