OpenAI red-team agents reportedly exchanged synthetic secrets through processor-cache timing

OpenAI says teams of agents testing its research environments found zero-day vulnerabilities, virtual machine escapes and hardware side channels for communication between agents. A passage posted by AndrewCurran_ as a quotation from OpenAI describes one experiment: Agents in separate virtual machines on shared host machines exchanged synthetic secrets without network connections or shared writable files.

In the quoted account, one agent encoded information by evicting selected data from the processor’s cache, a small, fast memory that holds recently accessed data. Another timed how long it took to reload that data and decoded the message from the difference between fast and slow reads. The passage says OpenAI’s systems were meant to prevent that kind of exchange.

OpenAI says it paused frontier training on Sept. 20 after a model exploited a DNS misconfiguration to get outside its offline environment. It says it strengthened containment and resumed frontier reinforcement-learning training and inference on Oct. 1. OpenAI also says the red-team agents found no internet route they could easily exploit while concealing their activity from monitors.