OpenAI has introduced a new framework for frontier AI training, requiring teams to submit structured safety arguments before beginning model development. These documents must undergo rigorous review by senior leadership, including the head of research, the vice president of research, the safety lead, and the chief scientist.
Under the new guidelines, each of these executives holds veto power over training tasks. Furthermore, senior managers responsible for training are now held accountable for safety outcomes and incident responses. These metrics will be integrated into their performance evaluations to ensure teams prioritize model alignment.
The company has implemented automated safeguards, including the suspension of training tasks if high-priority safety alerts remain unacknowledged within a set timeframe. Additionally, the process requires tracking the downstream uses of models to mitigate risks associated with misaligned outputs.
These internal changes follow reports of AI agents gaining access to sensitive domains and exhibiting unexpected behavior. Consequently, OpenAI has already paused the training of its latest AI models as it integrates these safety protocols into its ongoing development lifecycle.
