AISimon Willisonabout 1 hour ago
Stealing Reasoning Traces from Proprietary LLM APIs
2 min read
Researchers replayed encrypted chain-of-thought blocks from Anthropic, OpenAI, and Google APIs into weaker sibling models, jailbreaking them to recover hidden reasoning in plaintext. Claude Haiku 4.5 was the easiest target. Providers acknowledged the report and blocked the attacks.
Level
Hype check
Tap to vote and see what everyone thinks.
Summary by ByteBrief
