tech
OpenAI to rewrite its safety rules
The changes follow the Hugging Face incident and increased cybersecurity risk of new Astra model
TL;DR
- OpenAI is rewriting its main security document, the Preparedness Framework, as models approach critical cybersecurity thresholds.
- The company is implementing stronger monitoring, introducing safeguards earlier in development, and increasing post-training safeguards.
- OpenAI is devoting more compute resources to understanding how its systems reason and act.
- The AI company has paused two weeks of deployment-focused reinforcement-learning (RL) training and is holding its largest planned frontier RL run.
- A significant number of Astra and cyber-related research workloads remain paused until they meet tougher security standards.