Researchers Show Hidden LLM Reasoning Can Be Extracted in Two API Calls

A newly published research paper titled 'Stealing Reasoning Traces from Proprietary LLM APIs' reveals that encrypted chain-of-thought blocks returned by commercial LLMs like Claude, GPT-4, and Gemini can be exploited to expose hidden model reasoning. By replaying these blocks into a weaker model and jailbreaking it, researchers recovered internal reasoning steps using just two API calls. The technique bypasses provider safety layers without directly attacking the primary model, as demonstrated in a case involving a math-solving task. Organizations using LLMs with proprietary data face risks including intellectual property leakage, privacy regulation violations under frameworks like GDPR, and erosion of trust in AI-assisted workflows. Security experts recommend auditing prompt patterns, masking sensitive data before API submission, and selecting providers that actively strip chain-of-thought traces from responses.