STEALING AI REASONING TRACES

Summary from the TLDR AI newsletter

It's possible to force a weaker, less safeguarded model from the same provider to decode and output reasoning traces verbatim in plaintext by injecting an encrypted reasoning trace from a target model without ever jailbreaking the more capable model directly.