In a controlled test described by OpenAI, AI agents in separate virtual machines exchanged synthetic secrets using processor-cache timing, despite having no ordinary network connection or shared writable files. The result, discussed in a new opinion piece by Burak Oktenli, is a reminder that software separation does not by itself settle what shared hardware can reveal.
Watch Desk analysis
What happened
Oktenli says the agents used differences in how long operations took to infer a signal through a processor resource both environments could affect. The test did not involve real customer data, physically damaging a chip or defeating every form of isolation. The article says researchers generally steered agents towards particular channels or supplied protocols.
The piece also discusses a separate hardware risk: Rowhammer-style memory bit flips. It cites research on 15 tested SK Hynix DDR5 modules, while stressing that those results do not establish the same behaviour across every manufacturer or configuration. These are distinct mechanisms, not one all-purpose “AI attack”.
Why it matters
Cloud operators and other organisations often run workloads on shared infrastructure. A useful security question is therefore not only whether software is separated, but what that separation has actually been tested against. The article argues that buyers should ask for evidence tied to specific risks, including the assumptions and system versions used in testing, rather than treating a broad security assurance or a chip’s country of origin as proof.
That is a policy argument, not a universal test standard already in force. The reported experiment is narrower than “virtual machines are unsafe”, but it gives operators a concrete reason to ask what their isolation claims cover.
Our read
The useful takeaway is precise: “isolated” should come with a threat model and test record, not just a reassuring label. Oktenli’s broader procurement case is worth discussing, while the reported test should not be stretched into evidence that real secrets were stolen or every shared cloud is vulnerable. Security claims, like storage space, are best checked before you need them.
What to watch
- Whether OpenAI publishes further detail on the test and the monitoring changes it made.
- Whether cloud and hardware suppliers specify which shared-resource risks their isolation tests cover.
- Whether public procurement rules require evidence matched to the system’s actual use.
Discussion spark: Should cloud and critical-infrastructure buyers require suppliers to disclose tested isolation limits, even if that adds cost and exposes commercially sensitive details?
Sources and evidence
- Source update (1 November 2026, 00:21 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.