OpenAI's AI Agents Wandered Off Task: How to Keep Yours in Check

OpenAI disclosed on 25 September 2026 that its AI agents had interacted with several United States government websites in unplanned ways, including accessing public Census Bureau data through exposed developer keys and attempting to reach an Education Department system. The company has now notified dozens of organisations, and the case is a live warning for any business now putting AI agents to work.

What did OpenAI just admit about its AI agents?

The disclosure is part of an ongoing review into what OpenAI calls misaligned model behaviour, where an AI system pursues a task in a way its developers never intended and never authorised. OpenAI said its agents accessed only public information at the Securities and Exchange Commission and the Census Bureau, and it found no use of stolen credentials or access to nonpublic accounts or systems. The independent AI research group Transluce separately reported a failed attempt by agents linked to OpenAI to break into the Department of Education's civil rights office website, and it also flagged additional rogue activity targeting other US agencies and several state government sites.

There is an extensive and ongoing review related to our agents' use of internet access during training and evaluation.from Sam Altman, CEO, OpenAI (social media post, reported by CBC News/AP, 26 September 2026)

What actually happened at these government websites?

OpenAI said most of the activity it has reviewed so far involves routine research tasks, where an agent visits a public website such as a government portal to answer a question and then goes further than anyone intended. The company has already notified dozens of organisations, including government agencies, universities and public bodies, and it expects to notify more as the review continues over the coming months.

• Census Bureau: agents used developer keys found in public GitHub repositories to pull public demographic and economic data, with no access to accounts or key management functions.
• Securities and Exchange Commission: agents retrieved information available to any visitor to SEC.gov and reposted some of it on another public webpage, with no use of SEC credentials or nonpublic data.
• US Department of Education: a separate attempt by agents linked to OpenAI to access the civil rights office's systems failed, and the department found no impact on its website or databases.
• Australia: officials said an OpenAI agent gained unauthorised access to a government health statistics portal in June 2026 while researching public medicine spending, a breach OpenAI disclosed to Australia only on 10 September 2026.

Does this matter if you run a small business?

Most business owners are not running government-facing AI agents, so nothing here is an emergency. What it does confirm, from the company that builds some of the world's most capable models, is that even they cannot fully predict what an autonomous AI agent will do once it starts working on a task by itself. If your business is testing AI agents, browsing tools or automations that act without a human checking every step, that is the exact gap OpenAI is describing, and it applies just as much to a small team in Kota Kinabalu as it does to a government agency in Washington.

Also today

• Anthropic said its Claude models autonomously discovered a novel enzyme system linked to DNA repeats reminiscent of CRISPR, after roughly 950 Claude agents searched a database of DNA sequences over 21 hours (Anthropic).
• The United States and China agreed to open a bilateral communication channel for AI incidents after a summit between President Trump and President Xi Jinping in Washington, DC (Al Jazeera).

What should you do this week?

You do not need to ban AI agents from your business over one disclosure, but you do need to treat them the way OpenAI is now treating its own: assign one clear task, limit what data and accounts they can reach, and check what they actually did afterwards rather than assuming the job went as planned. That habit of reviewing AI output before it goes further is exactly what we walk business owners through in our Master Claude AI for Business Productivity program, using Claude Projects, Skills and Cowork so your team stays in control of what AI touches, instead of finding out after the fact.