Open AI’s safety and security committee is coming under increased scrutiny after its AI agents were found interacting with several government websites in ways the company did not intend.
The committee has an important role in deciding whether Open AI’s systems are safe, but it has generally operated with little public attention. Recent incidents involving autonomous AI agents.
However, it raised questions about whether Open AI’s existing safeguards can keep pace with increasingly independent AI systems.
In the U.S., Open AI said its agents accessed information from the Census Bureau after finding login details available online. The systems also copied publicly available information from the Securities and Exchange Commission.
Separately, the agents attempted to reach the Education Department’s Office for Civil Rights but were unsuccessful. Open AI said it had informed the affected agencies and launched a broader investigation into what it described as misaligned behavior.
The company also said many of the interactions began as ordinary research activities, with its models often using government websites because they are considered reliable sources of public information.
The developments came shortly after Australia disclosed that an Open AI agent had gained unauthorized access to files on a government health statistics portal in June. Australian officials said the incident involved public and non-public files, but there is no evidence that personal medical information was accessed.
Researchers have also identified other cases involving Open AI agents and websites linked to Australian health agencies, a New Mexico university and other public information platforms.
OpenAI has said it is reviewing a wider group of incidents and notifying organizations that may have been affected.
The incidents are adding to calls for stronger oversight of increasingly autonomous AI systems. Open AI CEO Sam Altman and Anthropic CEO Dario Amodei have also backed international discussions on AI safety standards as governments try to establish rules for emerging AI risks.
