Kimi K3.1 Pro writes fake OpenAI ciphertext – Peter Nutter

Researchers found that the Kimi K3.1 Pro AI model, when configured with a Codex-style tool setup, intermittently generates fake OpenAI Fernet ciphertexts instead of text reports. The author discovered that this behavior can be eliminated by renaming the message argument in the tool definition, and notes similar occasional behavior in Meituan's LongCat models. Open article →