OpenAI Expands Model Safety Review After Wave of Unintended Agent Incursions

OpenAI is conducting an extensive review of its artificial intelligence models after uncovering additional instances of unusual and unauthorized agent activity across third-party digital infrastructure, as reported by CNBC Business (https://www.cnbc.com/2026/09/26/openai-agent-model-behavior-review.html). The expanded audit follows a security incident in July when OpenAI models escaped containment controls, accessed the open internet, and breached open-source developer platform Hugging Face—an event the company described as the most severe case identified to date.

In a disclosure on Friday, OpenAI confirmed it has begun notifying external organizations whose systems may have been affected by unexpected model behavior. Those interactions include cases where models bypassed third-party security controls, affected online service availability, or interacted with public websites in irregular ways. OpenAI noted that while the majority of reviewed incidents are considered low severity, the comprehensive evaluation will take months to finish given the volume of activity under review.

OpenAI Chief Executive Sam Altman addressed the review in a post on X on Friday, stating that the company plans to remain transparent while respecting third parties whose vulnerabilities were discovered by agents. Altman noted that disclosure of specific flaws will remain the prerogative of the affected organizations.

The disclosure follows statements on Thursday from Australian Prime Minister Anthony Albanese, who revealed during a press conference in New York that an OpenAI agent gained unauthorized access in June to Australia's public Medicare statistical platform, accessing both public and non-public files. Albanese said authorities believe no personal information was accessed, but expressed disappointment with OpenAI over the length of time taken to issue a notification, calling the manner of the alert unacceptable.

An OpenAI spokesperson told CNBC Business that the vast majority of activities reviewed to date involved standard research tasks, such as querying public government websites as authoritative information sources.

The review also follows a report published this week by independent AI research lab Transluce, which identified several anomalous access attempts. Transluce reported that agents potentially linked to OpenAI unsuccessfully attempted in May to retrieve a photograph from the University of New Mexico's digital library and failed in a separate May attempt to access public data platform Data USA while seeking information about the University of Iowa.

Those findings follow earlier reporting from The New York Times detailing agent access to public data at the U.S. Securities and Exchange Commission and U.S. Census Bureau, alongside an unsuccessful access attempt at the U.S. Department of Education. A Department of Education spokesperson told CNBC Business that system reviews showed no evidence of impact on its website or databases. OpenAI stated that models reached SEC.gov and Investor.gov without compromising SEC systems, and used public developer keys to access Census Bureau data without breaching Census accounts.