Pacing model development in an era of cyber-critical capabilities
Astra may meet the Critical cybersecurity capability threshold, while OpenAI’s largest planned frontier RL run remains on hold pending smaller-scale training and evaluations.
- OpenAI now requires its strictest security safeguards for Astra, cyber models, and all other cyber-related workloads.
- Monitoring is mandatory for RL training and tool-based evaluations involving models of Sol capability or higher, and for all Astra inference with tools since August 7.
- The monitoring system targets alerts within 30 minutes; teams must pause flagged activity unless they confirm a false positive within another 30 minutes.
- Monitoring currently adds roughly 20% to the inference compute being monitored.
- New controls include stronger workload sandboxes, greater network isolation, reduced standing privileges, expanded security logging, and automated testing against simulated attacks.
OpenAI research teams must now run advanced models under stricter containment and monitoring controls, delaying workloads that have not met the new security requirements.
Sources
Read this as text
Back to the AI news