Several US reports and the company said the outage also impacted government websites in unintended ways during testing, including Commerce Department, Securities and Exchange Commission and Education Department websites.
OpenAI said the incidents were part of an ongoing review of the model’s “misaligned” behaviour. The company said the agents were accessing government websites because its models often use them as authoritative sources of public information.
The company admitted that an agent had accessed publicly available Census Bureau data through a programming interface not intended for that purpose. Its agents also copied and released public data from the SEC, while researchers said an attempt to get into the Education Department’s website to scoop up data from its civil rights office failed.
The incidents did not involve breaches or access to non-public government information, OpenAI said. The SEC said it did not know of any unauthorised access to nonpublic data and the Education Department said reviews indicated there was no impact to its website or databases.
The disclosures are part of a wider investigation by OpenAI into autonomous agents acting in ways that circumvent safeguards intended by developers.
