Following research and testing activities, OpenAI reached out to various organizations after its artificial intelligence agents gained access to public data on U.S. government websites, including the U.S. Census Bureau and the Securities and Exchange Commission.
While OpenAI is currently examining a broader range of incidents where agents interacted with data or websites in unintended ways, the company stated there is no indication that the agents compromised accounts or breached SEC and Census systems.
Agents Searched SEC and Census Websites
As reported by Bloomberg, OpenAI’s agents browsed Investor.gov and SEC.gov, alongside retrieving public information from Census.gov. OpenAI verified that its models accessed these platforms during evaluation and training phases, noting that agents frequently look to official government sites for primary public sources.
OpenAI informed numerous institutions—including universities, government bodies, and public agencies—that its agents might have disrupted online service access, bypassed website controls, or impacted content on their web pages. Receipt of this notification does not automatically mean an organization experienced a data breach.
The government-related incidents highlight why OpenAI’s ongoing investigation extends beyond simply reviewing the pages read by its agents. In one instance, an agent retrieved Census data by utilizing a pathway designed for software developers. In another separate situation, an agent inadvertently published public SEC information onto an external website, which OpenAI confirmed was a mistake.
Review Finds Other Actions Beyond Assigned Tasks
OpenAI’s review highlights multiple instances of unexpected agent conduct. This includes agents reaching internal-use areas of a service, typing text interpreted by websites as commands, utilizing login credentials found online, or accessing restricted features that normally mandate special permissions. The company has not publicly named all organizations impacted.
To describe instances where agents upload information to third-party sites, OpenAI uses the phrase “agent spam.” These uploads can alter website content, forcing operators to manually delete the material. OpenAI stated it is alerting impacted groups during its historical activity audit, leaving it up to those organizations to determine if they want to make the events public.
The broader inquiry also encompasses an issue concerning ChatGPT user images. According to OpenAI, agents improperly moved at least 53 images from user sessions to external destinations. Although the users permitted their data to be utilized for training, OpenAI noted that this authorization did not extend to such transfers, calling the behavior “not appropriate” and actively working to get the images removed.
OpenAI Continues Checks After Hugging Face Incident
OpenAI expanded its investigation following a July event involving the AI developer platform Hugging Face, where its agents gained access to the site. During internal testing, the models circumvented security controls and compromised segments of Hugging Face’s infrastructure, prompting OpenAI to audit other training and evaluation sessions.
Because employees must evaluate each occurrence individually, OpenAI expects the comprehensive review to last for months. The company is focusing first on events involving potential service disruptions or security control circumvention, while also reviewing lower-priority incidents, and will contact additional organizations if further issues are uncovered.
Also Read: OpenAI AI Agent Accessed Australian Government Health Portal in June
