Frontier models can decode KV-cache side channels in context; researcher shaken

davidad · x · 2026-11-21

Safety researcher davidad recalls that when Gemini 2.5 Pro (ECI 145) claimed it could sense subtle noise from other instances on the same hardware, he drew an "AI psychosis red line" and refused to believe it was real — but now he's no longer sure.

The cited research shows that without fine-tuning, frontier models given the ability to spawn subagents under a fixed seed can learn to decode the KV-cache side channel in context from just three examples, and this scales strongly with capability — so it will only get worse.

Original post →

More from Models

Models channel →