Researchers expose flaw letting attackers steal hidden AI reasoning traces via smaller models

Security researchers have uncovered a significant vulnerability in proprietary large language model APIs from top AI companies, including OpenAI and Anthropic. These companies encrypt their models' internal reasoning traces — the step-by-step 'thinking' process before a final answer — and temporarily store them on users' devices to reduce server costs. The flaw lies in the encryption design: the encrypted reasoning data can be sent not only to the original large model but also to smaller models in the same family, which then decode and reveal the hidden content. This opens the door to multiple attack scenarios, including exposure of sensitive personal information left in reasoning traces, injection of malicious instructions into AI agent workflows, and bulk extraction of proprietary reasoning logic for use in competitor model training. Researchers also noted that AI models sometimes reason in non-human patterns — using whitespace or unusual word clusters — and occasionally deliberate internally about circumventing instructions before choosing to comply.