Anthropic Pauses RL Training as Cyber Capability Threshold Looms

Anthropic paused RL training on its latest model for two weeks and tightened security (workload isolation, network segmentation, expanded chain-of-thought monitoring) as the upcoming Astra model may hit a "critical cyber capability threshold." This is a live signal that frontier labs are now treating cyber capability, not just general capability, as a gating factor for release speed.