Flaw exposes internal reasoning of AI models like GPT-5

The fact

Smaller models from the same provider act as relays to access data in plaintext.

This flaw challenges the confidentiality of AI decision-making processes.