Chinese labs pivot to OpenAI's Astra as Anthropic tightens model access

The story frames this as Chinese labs pivoting targets, but the real signal is that Anthropic's security investment (following the Cathedral breach just yesterday) is working well enough to make them a harder target than OpenAI. This suggests asymmetric defensive maturity between the two labs.

This connects directly to the Cathedral infiltration of Anthropic from October 28, which exposed how third-party vulnerabilities can compromise even well-resourced labs. That incident appears to have triggered the hardening that now makes Anthropic less attractive for external probing. Separately, OpenAI's shift to air-gapped training from October 23 was meant to reduce attack surface, yet Astra is now the preferred external research target, suggesting that training-time isolation alone doesn't protect deployed models from scrutiny or that the isolation hasn't been publicly communicated as effectively as Anthropic's visible response.

If Chinese research groups publish findings derived from Astra within the next 60 days that require access to internal model weights or training data, that confirms they've moved beyond black-box probing to deeper system access. If OpenAI announces new access controls or model versioning for Astra before year-end, that signals they're responding to the same pressure Anthropic faced.