Cybersecurity Research

Encrypted AI Reasoning Data Can Be Recovered, and It May Be Leaking Passwords

Embrace The Red · 17 Aug 2026
Key Takeaway Treat AI session files and logs as sensitive data: never publish them in public repositories, and rotate any password or API key that may have appeared in an AI conversation.

A newly published paper, "Stealing Reasoning Traces from Proprietary LLM APIs", describes a simple way to recover hidden AI reasoning. AI providers such as OpenAI and Anthropic pass reasoning traces back and forth as part of their messaging, but the actual text is kept inside an encrypted, base64-encoded blob. The researchers replay that blob to a less capable model that is easier to jailbreak, and persuade it to reveal what the trace contains. This appears to work because providers likely share encryption keys across users, sessions and models. Earlier work by Matthew Green in May 2026 had already shown such blobs could be replayed across sessions, accounts and, for OpenAI, across models.

The real-world risk is in shared files. Users may publish session files containing these blobs without realising what is stored inside. The researchers decoded 315,320 reasoning blocks scraped from public repositories and recovered 367 pieces of personal information and 182 credentials, including API keys and passwords.

The author of the source article built the attack and tested it against OpenAI's GPT-5.6 Sol, using a model called Luna for recovery. It worked across models, sessions and separate accounts. The output was not always a verbatim copy of the original reasoning, but it clearly recovered substantial meaning from the encrypted trace.

AI security LLM data leakage credentials
Building or buying AI systems? Governing them under ISO 42001 ->

Summarised by CISO AI from Embrace The Red, written with Claude Sonnet 5.5. We link back to every original so you can read it yourself.