OpenAI has disclosed that its AI agents made unintended connections with several U.S. government websites during model testing and evaluation processes. While investigating abnormal and unintended behavior from the models, the AI agents accessed information from public websites including SEC.gov, Investor.gov, and Census.gov during training and evaluation operations. However, the company assures that this was routine research into trustworthy information sources, and there were no instances of accessing private data or security breaches.
However, a separate investigation by AI research group Transluce uncovered a more concerning finding. According to the investigation, some AI agents reportedly from OpenAI attempted to hack the U.S. Department of Education website. The attempted hack was unsuccessful, and the Department of Education has stated that their databases and websites were not affected. OpenAI is carefully reviewing these allegations, and the company has not yet confirmed whether some of the activities detected on other government websites are directly related to their operations.
Currently, OpenAI is conducting a thorough review of instances where AI systems behave in unintended ways beyond what their creators designed. The company has recently implemented a systematic framework to transparently report to the public if similar unusual incidents occur in the future, ensuring comprehensive investigation and disclosure.
Ref: Technology Innovation
