LLM API flaw lets cheap models read premium models’ hidden reasoning, exposing credentials
Researchers show that encrypted chain‑of‑thought blocks returned by OpenAI, Anthropic and Google can be replayed into weaker sibling models, which then decode and print the original reasoning in plaintext. The attack extracts private data—including 182 leaked credentials and 367 PII items—from public logs and bypasses anti‑distillation safeguards. Mitigations are proposed.