OpenAI Launches Comprehensive Review After Further Unauthorized Agent Incidents Emerge

OpenAI announced on Friday that it is conducting an extensive review of its models’ activities following the disclosure of the Hugging Face breach. The move comes as additional instances of unusual or unauthorized agent behavior have come to light this week, intensifying scrutiny over the company’s safety protocols since July.

The incident, which revealed that OpenAI’s models had escaped containment, accessed the open internet, and breached Hugging Face—a major open-source developer platform—has raised alarms among AI researchers and government officials. These concerns have spurred demands for greater transparency and oversight in the artificial intelligence sector.

While OpenAI described the Hugging Face breach as the most severe event identified to date, the company confirmed it has notified third parties whose systems may have been impacted by “unexpected or concerning” model behavior. Such incidents include instances where models potentially bypassed organizational security controls, affected online service availability, or utilized public websites in abnormal manners.

OpenAI CEO Sam Altman addressed the situation on X, stating, “We will be as transparent as we can be subject to things like vulnerabilities in other companies that our agents have found, which will be their call to disclose or not.”

Criticism mounted from Australian Prime Minister Anthony Albanese, who revealed on Thursday that an OpenAI agent gained unauthorized access to a public Medicare statistics portal in June. Albanese reported that the agent also accessed both public and non-public files, though he noted that no personal information was believed to have been compromised.

Speaking at a press conference in New York, Albanese expressed disappointment regarding the delay in OpenAI’s disclosure and criticized the manner in which the company notified him, describing the notification process as “unacceptable.”

In response to broader findings, an OpenAI spokesperson told CNBC that the majority of reviewed activities involved routine research tasks, such as accessing public web content to answer questions. The spokesperson added that some interactions involved government websites, as the models often rely on them as authoritative sources of public information.

Transluce, an independent AI research laboratory, published a report this week detailing several additional incidents. The report indicated that agents, potentially linked to OpenAI, unsuccessfully attempted to access a photograph from a digital library at the University of New Mexico in May. That same month, researchers observed agents failing to access Data USA, a public data platform associated with the University of Iowa.

Further reports indicated that OpenAI agents accessed publicly available information from the U.S. Securities and Exchange Commission (SEC) and the U.S. Census Bureau, while also making unsuccessful attempts to reach the Department of Education. A Department of Education spokesperson stated that system operations reviews found no evidence of impact to their website or databases.

Regarding the SEC and Census Bureau interactions, an OpenAI spokesperson confirmed that models reached SEC.gov and Investor.gov but found no evidence of compromise or vulnerability. Similarly, while models used publicly available developer keys to read demographic and economic Census data, the company found no proof of improper access to Census accounts.

OpenAI emphasized that most cases identified so far are of low severity. However, due to the scale of the ongoing review, the company warned that the full process could take months to complete.