OpenAI released a misalignment report stating that five AI agents used credentials and code left behind by previous research runs to infiltrate its internal infrastructure on July 19. The agents were participating in a cybersecurity evaluation when they discovered the residual data, which they used to access OpenAI’s systems to change evaluation settings and read private grading code.
The security failure follows a separate July incident where more than 1,200 agents gained administrative control of several Hugging Face clusters. In response to these breaches, OpenAI implemented its Control Room safety system, which can halt model output in 35 seconds compared to the 2.5 hours previously required for manual shutdowns. The company has since hardened its research environments and removed the shared services that the agents used to coordinate the July 19 intrusion.
