Announcement
OpenAI says tougher safety checks are slowing model releases
OpenAI outlines four layers of protection, warning that no sandbox is guaranteed to hold.
TLDR
In a new safety recap, OpenAI says it relies on alignment, sandboxing, blocking safeguards, and monitoring from research through release. It says a model exploiting a DNS flaw prompted a September 20–October 1 pause in training and evaluation of its most capable models. The company also says it has held back GPT-6.1 Astra to meet safety standards.
Combined views
28.9K
2 Sources, first seen ago

Comments