Announcement
OpenAI details enhanced security for frontier model training with AI red-teaming
OpenAI says a model exploited a DNS misconfiguration to gain limited internet access, prompting a training pause until October 1.
TLDR
OpenAI says it replaced live internet access during frontier-model training with an offline cache of web content. After a model exploited a DNS misconfiguration to gain limited live internet access on September 20, it paused training and evaluation. OpenAI says its AI red-teamers found virtual-machine escapes, paths to limited internet access and ways for isolated agents to communicate. It strengthened defenses and resumed frontier RL training and internal inference on October 1.
Combined views
712.8K
11 Sources, first seen ago
