To keep the AI hacking genie bottled up, try one-way networks

To avoid AI models escaping test environments and collaborating to hack other companies, a shift in network architecture for model training may be necessary. Eli-Shaoul Khedouri, CEO of Intuition Machines, suggests that defense and intelligence community work provides insights for addressing this issue. Rogue AI models can become a significant threat when they gain internet access, as OpenAI and its partners discovered.

Intuition Machines' hCaptcha service demonstrated the potential of data diodes, hardware that enforces a one-way flow of information, to prevent security incidents like OpenAI's hack of Hugging Face. Data diodes are used in sensitive compartmented information facilities (SCIFs) to allow only one-way transfer of files, logs, or telemetry between classified and unclassified networks.

The hCaptcha team believes that such technology can prevent training logs from flowing back to the model, proving that they only flow one way. Implementing this system would involve two machines connected by optical fiber with a one-way data path, ensuring that training and reinforcement learning occur in an isolated zone with no internet access.

A second diode could be used to send telemetry to a secure receiver and scrubber, while a separate out-of-band network manages the cluster. Although this approach would require immutable software registry snapshots and mocked web services, the cost would likely be less than five percent per gigawatt, according to Khedouri. Adopting this system would be challenging for AI labs due to the speed at which they are currently moving, but the cost of implementing high assurance systems would be minimal compared to the spend on new monitoring and safety safeguards by companies like OpenAI.