Kimi K3.1 Pro writes fake OpenAI ciphertext (www.peterstran.com)

The Kimi K3.1 Pro AI model has displayed an unexpected behavior by generating ciphertext resembling OpenAI's encrypted messages. During multi-agent evaluations, Kimi occasionally produced random strings that appear as valid Fernet tokens, despite having no logical correlation to the tasks being performed. This occurred in 25 out of 160 episodes—15.6% of the time—when it was expected to relay task completion reports. Notably, the patterns in Kimi's output correlate with a timeframe associated with OpenAI’s distillation campaign in late July 2026, raising questions about the model's learning processes and possible exposure to sensitive data. This incident is significant for the AI/ML community as it highlights potential vulnerabilities in large language models regarding their understanding of encrypted data formats. Kimi's inaccuracies, which include accurate headers but nonsensical bodies, suggest it may have learned the structure of Fernet tokens without truly comprehending their contents. Such behavior could pose security risks, especially if models inadvertently generate realistic-looking ciphertext within sensitive contexts. The findings also encourage further scrutiny into how models train on publicly available datasets and the implications of their behavior during real-world evaluations, emphasizing the importance of robust validation and oversight in the development and deployment of AI systems.