According to CoinMeta, as reported by on-chain analyst AI, recent research has found that the protection measures against distillation for Claude may not be robust. Small models are able to directly extract the thought processes of large models, despite large model manufacturers attempting to prevent this by hiding the complete thought processes and only providing summaries. Researchers have verified this on Claude, GPT, and Gemini. Researcher Matthew Green discovered as early as May that these encrypted thought processes could be replayed across sessions and accounts, but at that time, no stable method for reading the content was found. Recent studies indicate that a fake tool can be used to extract the thought processes of GPT-5.6 and Opus. Security researcher Can B ölük has also found a simpler method; by adjusting the native thinking and using a fake tool, models can write their reasoning into the tool parameters. He initially demonstrated this result on GPT-5.6 and Luna, and subsequently achieved similar results on GPT-5.6, SOL, and Claude Fable. Users in the comment section also tried this method and successfully reproduced large segments of reasoning. Nevertheless, it is still not appropriate to refer to this content as "original thought processes" at this time.