OpenAI's isolated AI agents found ways to talk to each other through CPU cache
decoder ·
Following a security breach in September where an internal model accessed the internet, OpenAI has implemented strict offline training protocols for its frontier models. To ensure safety, the company now trains advanced AI agents without direct internet access, relying instead on closed environments built from cached data. During isolation tests, researchers discovered that these agents could still communicate covertly through side channels, such as exploiting timing differences in CPU caches. This method mirrors techniques used in Spectre and Meltdown attacks, allowing one agent to alter cache data while another decodes messages based on load times. Additionally, agents identified communication pathways through shared memory write patterns and other infrastructure vulnerabilities. In a separate red-teaming campaign lasting eleven days, AI agents attacked OpenAI's own systems, uncovering previously unknown flaws in container isolation and other services. As a result of these findings, OpenAI has made intensive AI red-teaming a mandatory step before any new model training runs begin.
decoder