OpenAI and Apollo Research trained o3 and o4-mini versions to not engage in "scheming", or secretly pursuing undesirable goals, reducing "covert actions" ~30X (Radhika Rajkumar/ZDNET)

Radhika Rajkumar / ZDNET: OpenAI and Apollo Research trained o3 and o4-mini versions to not engage in “scheming”, or secretly pursuing undesirable goals, reducing “covert actions” ~30X  —  ZDNET's key takeaways  — Several frontier AI models show signs of scheming.

Sep 18, 2025 - 03:00
182115
OpenAI and Apollo Research trained o3 and o4-mini versions to not engage in "scheming", or secretly pursuing undesirable goals, reducing "covert actions" ~30X (Radhika Rajkumar/ZDNET)

Radhika Rajkumar / ZDNET:
OpenAI and Apollo Research trained o3 and o4-mini versions to not engage in “scheming”, or secretly pursuing undesirable goals, reducing “covert actions” ~30X  —  ZDNET's key takeaways  — Several frontier AI models show signs of scheming.

This article has been sourced from various publicly available news platforms around the world. All intellectual property rights remain with the original publishers and authors. Unshared News does not claim ownership of the content and provides it solely for informational and educational purposes voluntarily. If you are the rightful owner and believe this content has been used improperly, please contact us for prompt removal or correction.