GPT-5 introduces a new safety-training method called safe-completion, which aims to balance helpfulness and safety by providing informative responses to dual-use prompts. This approach replaces the previous refusal-based training, allowing the model to navigate complex situations more effectively while adhering to safety guidelines.

OpenAI is preparing for advanced AI capabilities in biology by implementing safety measures and collaborating with experts to mitigate risks associated with dual-use technologies. This proactive approach includes training models to handle biological data responsibly while fostering partnerships to enhance biodefense research and policy development.