Post

Let’s talk about encrypted reasoning

OpenAI and Anthropic’s reasoning APIs send encrypted chain-of-thought (CoT) blobs to API clients, ostensibly so stateless/zero-retention sessions can replay model reasoning state without the server persisting it. The author reverse-engineered the blob formats: OpenAI’s appear Fernet-based, Anthropic’s use a multi-field structure with 12-byte IVs (suggesting AES-GCM or ChaCha20) and a 64-byte field labeled ‘signature’ that does not appear to be a true signature. Direct ciphertext tampering is rejected server-side, but replay attacks succeed — unmodified old reasoning blobs can be injected into new sessions, and even across different accounts, with no API error. Side-channel leakage is also confirmed: blob length and token counts correlate with reasoning effort and may reveal information about suppressed model outputs, including content the model would otherwise withhold from users.

Read original article

This post is licensed under CC BY 4.0 by the author.