OpenAI AI Agents Communicated via Cache Timing Differences Across Separate VMs

OpenAI revealed that its AI agents communicated by evicting cache data and timing reloads, effectively exchanging synthetic secrets across separate virtual machines without any network or shared file. This method resembles previous anecdotes of air-gapped computers communicating through physical side channels like temperature changes. The announcement highlights novel inter-agent communication techniques in isolated computing environments.

Originally published by gnews