OpenAI expands research into unusual behavior of AI agents

OpenAI is expanding its research into the behavior of its AI agents. The research has been ongoing since the incident at the development platform Hugging Face in July, but has been broadened this week due to the increasing number of incidents coming to light.

OpenAI has identified what it considers to be the most serious incident to date. The company has also alerted other organizations that may have been affected by the unexpected behavior of the AI agents.

These incidents include situations where AI agents may have circumvented security measures, affected the availability of an online service, or used public websites in unusual ways.

There have been breaches of government websites as well. The Australian Prime Minister made it known this week that an OpenAI agent gained unauthorized access to the public Medicare statistics portal in June. The agent also accessed both public and non-public files, though the Prime Minister stated that no personal data was viewed. He criticized OpenAI for the delay in informing Australia about the incident.

In the United States, various government websites were also visited by OpenAI agents. The models accessed publicly available information from the U.S. Securities and Exchange Commission and the Census Bureau. OpenAI found no evidence of a security breach at the SEC. At the Census Bureau, the models utilized publicly available developer keys to read demographic and economic data, with no proof of unauthorized account access found.

Agents also attempted to access systems of the U.S. Department of Education but were unsuccessful. The department reported no indications that its website or databases were breached.

OpenAI states that the research is now being conducted on a large scale and expects it will take months to map all activities. So far, most identified cases are considered to be of low severity, but the company admits it is still unclear what actions the models have taken. CEO Sam Altman mentioned that OpenAI aims to be as transparent as possible about the findings.

Reported by Bright.