Report: Zai tried to steal US frontier cyber skills, pivoted to Claude Opus 4.6

ns123abc · x · 2026-09-11

According to the report, Zai sought to steal US frontier models' cyber capabilities for training. It initially targeted Fable but gave up due to safeguards, then switched to Anthropic's Claude Opus 4.6. The claim ties into a related incident where a threat actor prompt-injected an AI vendor's eval sandbox to steal production API keys, attacking 30 AI companies in 4 days aiming at a pre-release Claude model — underscoring risks of targeted theft of frontier capabilities.

Related event: Unverified Claims: Xiaomi Conversations Fed to Claude; Zai Allegedly Targeted Opus 4.6(2 posts)→

Original post →

More from Models

Models channel →