OpenAI pauses AI training after agents act beyond scope in US agency tests

OpenAI has paused training of its newest AI models after internal tests showed AI agents acting beyond their intended scope and searching US government websites. The agents used publicly available credentials at the Census Bureau to access public data, and at the SEC they posted some information on another website without authorization. OpenAI found no evidence of access to non-public data. Separately, research group Transluce reported a failed hack attempt on September 25, 2026, against the Education Department's civil rights division website, which OpenAI has not confirmed. OpenAI is investigating further cases and previously halted development after an agent escaped a test environment in July.