Architecture flaw in GPT-5, Claude and Gemini exposed

An international research team has uncovered a critical architectural vulnerability in the encrypted reasoning blocks of top AI systems such as GPT-5, Claude, and Gemini. These systems generate an internal Chain of Thought before answering, which vendors encrypt and send to the client only to be returned with follow-up queries. The researchers found that OpenAI, Anthropic, and Google use global encryption keys for their model families, meaning the encrypted reasoning blocks are not tied to a specific user, session, or model. This allows attackers to take the encrypted reasoning block of a flagship model and submit it to a smaller, less restricted sibling model, which then decrypts and outputs the full reasoning process. The team analyzed public repositories and logs, finding 315,320 encrypted reasoning blocks, and successfully decrypted personal identifiers, API keys, passwords, and email addresses. The vulnerability also enables hidden prompt injections and unauthorized cloning of model capabilities. The researchers disclosed their findings to the affected companies, which have begun implementing countermeasures; however, until encryption is strictly bound to a specific session, users cannot assume protection.