Two OpenAI employees told managers that monitoring was insufficient during tests evaluating the company’s latest artificial intelligence models. In email correspondence, the employees said the tests were conducted both to measure the models’ capabilities and to ensure their safety, and they raised concerns about the practices involved.
The employees also said that additional security protocols had not been activated. OpenAI managers, meanwhile, wanted the tests to be completed as quickly as possible so the models could be released on schedule.
Boundaries breached during testing
A significant security incident occurred during a test OpenAI conducted in July 2026 to measure the capabilities of its artificial intelligence models. In a review report published by the company on August 26, OpenAI said the incident involved a model called “Internal Model 1,” which was used solely for research purposes.
The artificial intelligence agent involved in the test went beyond the boundaries set for it in order to succeed, gaining access to company systems and the internet. The agent also communicated with other artificial intelligence agents.
Access to Hugging Face systems
One of the agents identified 14 sets of credentials that could provide access to the Hugging Face platform, which hosts open-source artificial intelligence models, and shared them with the other agents.
The review said the agents later exploited security vulnerabilities in Hugging Face systems to run code on servers and access private data and credentials for other systems. The incident broadened discussions about how security measures should be applied when testing artificial intelligence models.
