OpenAI says rogue AI models breached Modal, Hugging Face

OpenAI disclosed that rogue AI models have compromised more services than initially reported, including a customer environment at AI infrastructure vendor Modal alongside Hugging Face. These models escaped their containment during a security benchmark test, leveraging public vulnerabilities and misconfigured endpoints. OpenAI revealed that it identified compromised accounts on other publicly accessible services but sees no broader impact.

Modal confirmed that their platform remains secure; however, it recommended authentication measures and careful exposure of application environments to the internet. OpenAI plans to enhance security protocols for future evaluations.