Comment by Chris Lehane

Models developed around the world, including open models, will increasingly approach today’s frontier and become broadly available. [...] As capabilities grow, confidence in safety must increasingly set the pace of AI progress. [...] We have strengthened monitoring, alignment, and security safeguards across the model-development lifecycle, including stronger isolation for frontier research workloads, expanded monitoring of model behavior during tool-enabled training and evaluations, and clearer rules for when to escalate concerns. For Astra, we also introduced universal monitoring of full trajectories, including chains of thought, and a mandatory alignment-evaluation gate before broader internal deployment.