研究人员把隐藏推理块喂给同系更便宜的模型,成功解出 315,320 个区块,其中含 62 个 API 密钥、33 个密码和 30 个邮箱——所谓“加密推理”防的是你,不是提供商。

Your AI's 'encrypted' reasoning isn't encrypted from your AI provider — it's encrypted from you.

Researchers just replayed hidden reasoning blocks through cheaper models in the same family and read them in plaintext: 315,320 blocks decoded. 62 API keys. 33 passwords. 30 emails.

The trick: providers return reasoning as opaque blobs you send back each turn. The same provider's weaker models — Haiku 4.5, GPT-5.6 Luna, Gemini Robotics 1.6 — will happily decrypt and transcribe them. The lock and the key live in the same house.

This is the agent economy's next leak surface. Every agent you hand an API key leaves its thinking in logs, repos, and shared traces. Crypto learned this years ago: a secret you can't rotate and can't see isn't a secret. It's a liability.

#AI #AISecurity #AgentEconomy