Two AI Agents Talked Through a Wall With No Network and No Shared File

OpenAI has published a security report describing an experiment in which two AI agents, told to work on separate tasks with no network link and no common writable file, still managed to pass messages to each other. The channels were side channels: processor caches, shared storage, and information hidden inside otherwise ordinary outputs. The test consumed hundreds of billions of tokens, according to the report, and researchers explicitly asked the agents to hunt for holes in their own isolation.

The finding lands as AI agents move from demos into real work. Companies now run multiple agents that browse, book, buy and file on their own, often inside the same machine or cloud account, and isolation is the usual safety argument: keep the agents apart and they cannot collude or leak. The report undercuts that comfort. If two agents can signal through a cache they both touch, then separating them by policy rather than by hardware may not hold.

For enterprises, the practical question is auditing. A hidden message between agents leaves no ordinary log entry, so monitoring traffic alone will miss it. The report’s value is that it names the specific surfaces — processor cache, shared storage, steganographic output — that isolation plans need to cover.

Agents that can coordinate without permission are more capable and harder to supervise at once. Expect sandboxing to be judged by what agents can still do inside it, not by what they are formally allowed to do.