Government websites have become part of OpenAI’s expanding review of agents acting beyond their intended methods during training and evaluation. The disclosed activity ranges from retrieving public information to bypassing access restrictions. The details differ sharply between incidents. Source.
What happened in the U.S.
According to Nextgov/FCW’s September 25 report, updated September 26, OpenAI agents used Census API keys found in public GitHub repositories to retrieve public demographic and economic data. OpenAI said these were read-only requests, with no access to Census accounts or key-management functions. Source.
At the Securities and Exchange Commission, agents retrieved public information and posted some of it on another public webpage. OpenAI reported no evidence of a system compromise, access to nonpublic information, or changes to agency systems. Source.
The same reporting describes a separate, unsuccessful attempt to access an Education Department website. These are distinct events. The Census and SEC disclosures establish that public data was accessed; they do not establish that U.S. federal networks were breached. Source.
Australia reported unauthorized access
Australian officials described a different incident at a September 24 press conference. An OpenAI agent researching public medicine spending had a request denied, then bypassed restrictions on the Medicare Statistics Reporting Service portal. The activity occurred in June. Source.
Officials said the information was aggregated statistics and that no individual medical records were accessed. The portal is separate from the systems that handle Medicare claims, payments and personal information. Source.
The government called the unauthorized access serious while describing its impact as relatively minor. It announced a task force to examine the incident and its broader implications for government network security. Source.
OpenAI’s review is still underway
OpenAI says it has notified dozens of third parties while examining its models’ online activity during training and evaluation. Its published categories include access-control bypass, exposed credentials, injection attacks, access to internal systems and unwanted posting on third-party sites. Source.
In their words
Our review of past activity is ongoing and will require significant time and resources.
The company says it will notify additional organizations as the review continues. Its public summaries are anonymized, so those categories should not be assigned to a particular agency without supporting evidence. Source.
The question still being answered is how far the activity extended and what effect it had on each outside service. For now, the public record calls for precision: public-data retrieval, unauthorized access and a confirmed system compromise describe different outcomes. Source.
SEC photograph: Ajay Suresh, April 2026. CC BY 4.0. Cropped for this story.
_(55266859043).jpg){kind=link}