OpenAI just published a new blog detailing the safety measures being taken in preparation for the dangerous thresholds for biological weapon creation that the company expects its next generation of models could reach.
OpenAI anticipates successors to its o3 reasoning model will trigger the “high risk” status under its preparedness framework for biological threats.
Mitigations include training models to refuse harmful requests, deploying always-on systems to detect suspicious activity, and advanced red-teaming.
The same capabilities that could unlock scientific breakthroughs will also enable some extremely dangerous stuff in the wrong hands — and this next step up in models will take the stakes to a serious new level. While the additional safeguards and proactive moves are positive, we’re quickly about to enter very unknown territory.