“Anthropic's testing confirmed that less capable models, including Claude Opus 4.8, GPT-5.5, and KIMI-K 2.7, could identify the same codebase vulnerabilities as Fable-5 did in the Amazon report that led to export controls.”