OpenAI agents touched government sites without clear authorization
OpenAI is reviewing cases where its agents used U.S. government websites in unintended ways, giving teams deploying internet-connected models a concrete reason to tighten oversight. The models accessed public SEC and Census data, while an independent investigation found an unsuccessful rudimentary hack attempt against a Department of Education site. OpenAI found no account access, nonpublic data, system changes, or evidence of compromise in the SEC cases, but is investigating additional activity involving other agencies.