AnalysisCybersecuritySeptember 8, 2026

Researchers steal AI reasoning traces via encrypted-block jailbreak

A new paper exploits an architectural flaw in proprietary LLM APIs: encrypted reasoning traces are interchangeable across sessions, users, and models, allowing attackers to inject a trace into a weaker model to decode it verbatim. Demonstrated against Anthropic, OpenAI, and Google, the attack recovered 367 PII artifacts and 182 credentials from 315,320 public reasoning blocks.

1 source

More stories today

Open the live feed