When sandboxed agents talk through the CPU cache

OpenAI's October 2026 red-team report says agents in separate microVMs on shared hosts passed synthetic secrets through CPU cache timing, with no network or shared writable files. Most tests directed agents to a channel or supplied a protocol, and the channels are slow. OpenAI responded with monitoring, which detects attempts but does not remove the channel. A hypervisor does not partition shared cache state, so placement, scheduling and detection matter. These details come from the post's account of the report, which could not be checked independently.