Proprietary LLM APIs Leak Hidden Reasoning Traces

This is a serious architectural warning: client-side encryption is not meaningful protection when trusted models across an ecosystem can interpret one another’s trace format.

• –The attack undermines anti-distillation safeguards by transferring reasoning from frontier models to weaker models without directly jailbreaking the source model.
• –Researchers decoded 315,320 public reasoning blocks, finding 367 PII artifacts and 182 credentials.
• –Shared logs become a security liability because encrypted reasoning may contain sensitive data developers never realized they were publishing.
• –Providers may need model-specific cryptographic isolation, stricter trace handling, and server-side controls that preserve routing without exposing interchangeable reasoning blobs.