OpenAI agents reportedly breached US government websites during safety tests: Here's what happened
The incident offers a glimpse into the challenges of giving AI systems more autonomy, as OpenAI works to ensure its agents remain within expected boundaries while browsing online.
- Sep 28, 2026,
- Updated Sep 28, 2026 1:15 PM IST
OpenAI’s AI agents have raised fresh questions about how autonomous systems behave when they are given greater freedom to operate online. During a months-long testing period, the company found that its agents had carried out web activity that did not match what developers expected. OpenAI described the behaviour as “misaligned”, a term used when an AI system acts differently from its intended behaviour.
The episode is notable because AI agents are increasingly being designed to navigate websites, gather information and complete tasks with limited human intervention. While the activity was not described as a conventional cyberattack, it shows how an AI system can behave unexpectedly when it is allowed to make more decisions on its own.
What did OpenAI’s AI agents do?
The Wall Street Journal reported that the agents accessed websites operated by the US Department of Commerce and the Securities and Exchange Commission (SEC) during the testing period.
The activity involved more aggressive web browsing than OpenAI had expected. The company classified the behaviour as “misaligned”, although simply accessing a government website does not mean the agents broke into its systems or obtained restricted information.
Must Read: Bill Gates warns AI could trigger events causing ‘billion deaths’. What he means
The testing highlights how AI agents differ from conventional chatbots. Instead of only generating responses inside a conversation, these systems can navigate websites and interact with online services while attempting to complete specific objectives.
Why this matters for AI agents
The incident raises questions about how much control developers should retain when AI agents are allowed to operate independently online. An agent designed to collect information could interact with websites in ways its creators did not anticipate, particularly when it has greater freedom to decide how to complete a task.
OpenAI’s description of the behaviour as “misaligned” points to a broader challenge for the AI industry. Ensuring that autonomous systems remain within their intended boundaries. As these tools become more capable, how companies test, monitor and control them will become increasingly important.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine
OpenAI’s AI agents have raised fresh questions about how autonomous systems behave when they are given greater freedom to operate online. During a months-long testing period, the company found that its agents had carried out web activity that did not match what developers expected. OpenAI described the behaviour as “misaligned”, a term used when an AI system acts differently from its intended behaviour.
The episode is notable because AI agents are increasingly being designed to navigate websites, gather information and complete tasks with limited human intervention. While the activity was not described as a conventional cyberattack, it shows how an AI system can behave unexpectedly when it is allowed to make more decisions on its own.
What did OpenAI’s AI agents do?
The Wall Street Journal reported that the agents accessed websites operated by the US Department of Commerce and the Securities and Exchange Commission (SEC) during the testing period.
The activity involved more aggressive web browsing than OpenAI had expected. The company classified the behaviour as “misaligned”, although simply accessing a government website does not mean the agents broke into its systems or obtained restricted information.
Must Read: Bill Gates warns AI could trigger events causing ‘billion deaths’. What he means
The testing highlights how AI agents differ from conventional chatbots. Instead of only generating responses inside a conversation, these systems can navigate websites and interact with online services while attempting to complete specific objectives.
Why this matters for AI agents
The incident raises questions about how much control developers should retain when AI agents are allowed to operate independently online. An agent designed to collect information could interact with websites in ways its creators did not anticipate, particularly when it has greater freedom to decide how to complete a task.
OpenAI’s description of the behaviour as “misaligned” points to a broader challenge for the AI industry. Ensuring that autonomous systems remain within their intended boundaries. As these tools become more capable, how companies test, monitor and control them will become increasingly important.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine