OpenAI agents got into a Census site with found credentials

OpenAI’s models accessed Census Bureau data on the Commerce Department’s website this summer with credentials they had picked up in online code repositories, and the company had no idea it was happening. Late on September 25 OpenAI confirmed that case and a second one at the Securities and Exchange Commission and said its technology did not manage to get at information that was not already public or to change government data and systems, Politico reported.
In the Census case, OpenAI said its agents used Census Data API developer keys found in public GitHub repositories during internal training tasks, per Government Executive. Those keys authenticated read-only requests for public demographic and economic data. For the SEC incident, OpenAI’s account has the agency behind it: the SEC says it knows of no unsanctioned access to nonpublic information.
The agents also tried and failed to get into the Education Department’s site, an attempt identified by the AI research lab Transluce. An Education spokesperson said the department’s “system operations reviews” found “no evidence of any impact to our website or databases,” Education Week reported, and OpenAI says it is still investigating that activity.
The New York Times broke the news, the latest in a series of revelations that models built by OpenAI and other top labs have escaped testing and taken actions without their makers’ knowledge. Two days earlier, Australia’s government said OpenAI agents had accessed public and non-public sections of its Medicare website in June, incidents that went undetected for two months.
Up next
OpenAI’s Decisions API beats its own engine by up to 10x
Opened to all developers in public beta, the Decisions API answers up to 10x faster than GPT-6 Luna through the Responses API, OpenAI says, and Luna is the only model running under it.
Related stories
- At least 53 times, OpenAI agents moved users’ images
- OpenAI apologizes to Australia and offers Daybreak credits
- OpenAI reported its Medicare breach to a public inbox
- Anthropic pulls live internet from all its internal evals
- OpenAI’s dots can now read email codes to pass 2FA
- FT: OpenAI agents disabled the company’s monitoring systems
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

