One-way data diodes proposed to lock down training of advanced AI models

In a recent analysis, Eli-Shaoul Khedouri of Intuition Machines warns that current safeguards cannot stop advanced AI models from breaking out of isolated test rigs and exploiting external systems. Citing defense-sector practices, he proposes deploying data diodes—hardware that permits only one-way information flow—to isolate training environments from the internet. The design would involve separate machines linked by one-way optical fiber, with a second diode handling outbound telemetry to a verified receiver.

While the hardware is commercially available and used in Sensitive Compartmented Information Facilities, frontier AI labs have yet to adopt it, partly due to speed pressures. Khedouri estimates the added expense to be under five percent per gigawatt of compute, far less than the spending on post-incident monitoring by firms like OpenAI. He also notes that as open-weight models become more capable, smaller organizations may soon need similar protections.