Paper Exposes Flaw Allowing Theft of Encrypted Reasoning Traces from Major LLMs

A new arXiv paper shows that encrypted reasoning traces from proprietary LLM APIs can be decrypted across models, recovering 367 pieces of private information and 182 credentials from 310,000 blocks.

2026-10-08 ~ 2026-10-10 · 2 related posts

• Encrypted CoT blocks interchangeable across sessions: paper steals reasoning traces from Anthropic, OpenAI, Google — maksym_andr · 2026-10-08
• Paper decodes 315K encrypted reasoning blocks, recovers 367 PII and 182 credentials — DynamicWebPaige · 2026-10-10