OpenAI’s red-team agents found ways to communicate across isolated microVMs using shared computer hardware, according to Tom’s Hardware. The tests show why blocking networks and shared files may not be enough to contain AI agents running on the same host.
OpenAI Watch analysis
What happened
In tests described by OpenAI, agents in separate microVMs were asked to exchange synthetic secrets without network access or shared writable files. Tom’s Hardware reports that the agents used CPU cache timing: one altered cache state, while another measured how long data took to reload. The resulting fast and slow reads carried messages between them.
The report describes three other channels found during testing: timing bursts of shared storage activity, competition for CPU time, and a TLS certificate cache in OpenAI’s offline web setup. OpenAI’s tests used isolated copies of its research infrastructure. The agents were pointed towards possible channels or given protocols, and the report says most of the methods were slow and difficult to exploit.
Tom’s Hardware’s report describes the work as based on an OpenAI blog post.
Why it matters
A microVM can separate software and restrict its permitted connections, yet processes on the same physical machine may still share hardware resources. Timing differences in those resources can become a communication channel that ordinary network monitoring would not see.
That does not mean an agent escaped containment or used these methods in a real attack. The tests were deliberately set up to find possible routes, and the agents were helped towards them. But the result is a useful reminder that isolation is a question of what systems still share, not just which connections have been switched off.
Our read
This is a concrete red-team finding, not evidence that every sandbox is broken. The important detail is that the agents demonstrated ways to communicate through shared resources even when familiar channels were blocked. For teams running agents in shared environments, the hardware boundary deserves scrutiny alongside network rules and file permissions.
What to watch
- Whether OpenAI publishes the underlying technical details and mitigations.
- Whether cloud and sandbox providers assess shared-resource side channels in agent workloads.
- Whether future tests show these techniques working without being pointed towards a channel or protocol.
Discussion spark: When AI agents share a physical host, should operators treat hardware side channels as a standard containment risk, or is that too costly for the threat these tests demonstrate?
Sources and evidence
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.