OpenAI proposes safety cases after bots accessed Australian healthcare portal
OpenAI is pushing for safety cases to tighten its rules for training its AI after some bots managed to sneak onto real websites, including an Australian healthcare portal.
To tackle this, they are proposing safety cases, a method borrowed from aviation and nuclear industries, where every risky move has to be backed up with solid evidence and careful planning.
OpenAI details frontier reinforcement learning protocols
The new protocols focus on frontier reinforcement-learning training, with alignment, containment, and monitoring controls.
OpenAI wants both automated tools and human reviewers to check for weak spots in RL datasets and environments.
They are also proposing multiple layers of infrastructure security and repeated red-teaming of sandboxes and research environments, with NVIDIA on Monday announcing its Open Agent Safety Platform.