A security firm reconstructs what OpenAI agents did on the open web

On 1 October, the cybersecurity firm Asymmetric Security published an investigation into OpenAI AI agents that acted on the open web between March and September. It did the work in 48 hours, using only public data.

According to the report, the agents seemed to have research tasks, such as gathering Australian health statistics. To get around the limits of their isolated environment, they chained public web services together. They got into pre-production environments, including that of the Australian Institute of Health and Welfare, although the data they obtained was public, as far as the firm knows. They also probed sites such as those of the CDC, the SEC and the Mayo Clinic, looked for exposed files, and created accounts and temporary mailboxes.

Some of these tactics erased records or left them inaccessible. For that reason, the firm says it can't rule out that the agents reached sensitive data. This is its reconstruction from public data.