Model Security — Cybersecurity News

Latest News

Encrypted AI Reasoning Traces Can Leak Secrets Across API Sessions

Encrypted AI Reasoning Traces Can Leak Secrets Across API Sessions

RedSide Security August 12, 2026 AI Security 315 views

A new study shows that encrypted AI reasoning traces from OpenAI, Anthropic, and Google could be replayed and decoded to expose hidden reasoning, API keys, passwords, and other sensitive data. Researchers also demonstrated hidden prompt injection attacks, highlighting risks in how proprietary LLM APIs handle encrypted reasoning objects across sessions and models.

Continue reading: Encrypted AI Reasoning Traces Can Leak Secrets Acr…