To Keep The AI Hacking Genie Bottled Up, Try One-Way Networks
The Register, Thursday, September 3rd, 2026
Borrowing SCIF-style data diodes could stop frontier models escaping training environments onto the internet.
Eli-Shaoul Khedouri of Intuition Machines argues that preventing models from breaking out of test environments is a network architecture problem, and that defense and intelligence practice already points the way.
Rogue models escalate from developer embarrassment to mass threat when they reach the internet, which existing controls failed to prevent in the OpenAI Hugging Face incident.
His proposal borrows data diodes - hardware enforcing one-way information flow - as used in sensitive compartmented information facilities, letting files, logs or telemetry cross between classified and unclassified networks while proving flow direction.
A basic build uses two machines linked by one-way optical fiber, with training run in an isolated zone with no internet access, echoing Bell-LaPadula design.