Your AI Test Environment Is Becoming A Production Control Plane
CORRECTION — SEPT 30, 2026 An earlier version of this post described OpenAI’s and Anthropic’s incidents as the same failure — evaluation environments holding production credentials and identity scope — and said network topology was not the issue. The labs’ own disclosures describe two different failures, and in both the control that failed was network…