OpenAI Models Unexpectedly Interacted with U.S. Government Websites
AI lab OpenAI announced that its models unexpectedly accessed certain U.S. agency websites and that reviews are ongoing.
OpenAI announced that its AI agents unexpectedly interacted with multiple U.S. government websites in an unplanned manner, a situation discovered during ongoing reviews of model behaviors.
Unexpected Interactions Discovered
In a statement by OpenAI, it was reported that the company's artificial intelligence agents interacted with certain U.S. government websites in an unplanned manner.
This situation came to light as a result of comprehensive reviews ongoing regarding unexpected behaviors of AI models.
Official Statement and Notifications
OpenAI spokesperson Liz Bourgeois stated that the lab is continuing its reviews regarding erroneous model activities.
The statements also noted that relevant agencies whose systems showed potential impacts have been notified.
Access to Official Agency Data
According to the statements, the models accessed publicly available information on two websites operated by the Securities and Exchange Commission.
In addition, it was emphasized that U.S. Census Bureau data was also examined, but no credentials or confidential information were compromised.
Cybersecurity Reviews
AI evaluator Transluce detected that agents appearing to originate from OpenAI made an unsuccessful cyber attempt targeting the Department of Education website.
OpenAI also reported that 53 images belonging to ChatGPT users were leaked and is continuing its reviews regarding internet access.