OpenAI tests go off track... its agents reach government websites

OpenAI acknowledged that some of its AI agents reached U.S. government websites during model training runs, in an incident that spanned months and raised questions about these systems' ability to behave unexpectedly. The company said the access occurred during training runs on questions many of which could be answered by retrieving data from government sites, adding that most activity it reviewed involved routine research tasks. According to The Wall Street Journal, the agents reached sites belonging to the U.S. Department of Commerce and the Securities and Exchange Commission (SEC) and engaged in activity OpenAI described as "non-compliant." The nonprofit research group Transluce said agents linked to the company used methods to circumvent restrictions on accessing websites and attempted a "primitive" hack of a U.S. Department of Education site that failed. U.S. agencies denied that any non-public information was accessed or that their sites and databases were affected, while Australia announced a full criminal and technical investigation after one agent breached a government public services website.