One of the most heated debates in the artificial intelligence industry has developed unexpectedly after the publication of a new scientific paper. Experts are trying to find out whether the Chinese company Moonshot could have created its Kimi K3 model by distilling it from the advanced Claude Opus neural network from Anthropic. Critics have previously pointed out that the time interval between model launches is too short, arguing that large-scale copying is almost impossible in such a period. However, a newly discovered vulnerability in software interfaces may be the very clue that American laboratories have been looking for for a long time.

Modern AI models form internal chains of reasoning before providing a response, which are usually locked inside encrypted text blocks. New scientific work has revealed a serious problem in this protection scheme. It turned out that the largest providers, including OpenAI, Anthropic and Google, used global encryption keys throughout their ecosystem. Moreover, the encrypted block of thoughts of a powerful model can be copied and submitted to a weaker related version, which lacks strict constraints and is able to decipher hidden thoughtsModern AI models form internal chains of reasoning before providing a response, which are usually locked inside encrypted text blocks. New scientific work has revealed a serious problem in this protection scheme. It turned out that the largest providers, including OpenAI, Anthropic and Google, used global encryptions throughout their ecosystem. Moreover, the encrypted block of thoughts of a powerful model can be copied and submitted to a weaker related version, which lacks strict constraints and is able to decipher hidden thoughts. This creates three serious threats at once: leakage of personal data and secret credentials, circumvention of built-in security measures, and the possibility of hidden malicious injections. The researchers decrypted public logs and found hundreds of pieces of confidential information, including passwords and access keys.

As an experiment, the authors submitted only one percent of the deciphered Claude Opus arguments to Kimi K3, after which the thinking and responses of the Chinese model sexperiment, the authors submitted only one percent of the deciphered Claude Opus arguments to Kimi K3, after which the thinking and responses of the Chinese model shifted sharply towards the Claude style. Statistical analysis has shown that specific segments of Claude and GPT reasoning are extracted from Kimi K3 about six orders of magnitude easier than from other open models. This indicates a strong behavioral compatibility with the reasoning patterns of the American neural network. It is unlikely to be possible to definitively prove the fact of distil