Crypto
Home›Crypto›Market Structure›Researchers decode encrypted AI reasoning logs and rec…
Researchers decode encrypted AI reasoning logs and recover live credentials
The study said researchers decoded 315,320 encrypted reasoning blocks and found 62 live API keys and 33 passwords in public session logs.
Researchers say they found a security weakness affecting major AI reasoning models used by providers including Anthropic, OpenAI, and Google, by decoding encrypted “reasoning tokens” that are shared back to providers in follow-up messages.
According to Decrypt, the researchers scraped 315,320 encrypted reasoning blocks from public GitHub and Hugging Face repositories, and through decoding recovered 182 credentials and additional personal information artifacts.
The paper, submitted August 10 by researchers and organizations including MATS Research, the ELLIS Institute Tübingen, the Max Planck Institute for Intelligent Systems, and security firm Snyk, said the flaw is architectural, with providers using a single global encryption key rather than tying encrypted blocks to specific users, sessions, or models.
Decrypt reported that after responsible disclosure, OpenAI, Anthropic, and Google deployed server-side patches, but historical session logs that had already been shared publicly remain decodable.