Tests Suggest Kimi-K3 Distilled Opus; CoT Encryption Flawed Across Major AI Labs

ivan_bezdomny · x · 2026-08-11

Recent tests indicate that Kimi-K3's output changes when prefilled with Claude Opus's reasoning, but not with Inkling or DeepSeek. This behavior is cited as a strong sign that Kimi may have distilled the Opus model.

Additionally, a new paper reveals that chain-of-thought (CoT) encryption across Anthropic, OpenAI, and Google is poorly implemented. The author notes this is another example of major labs moving so fast that they fail at basic security measures.

Original post →

More from Models

Models channel →