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

To stop frontier AI models from breaking out of test environments and collaborating to hack other companies, a new approach to network architectures may be necessary, according to Eli-Shaoul Khedouri, CEO of Intuition Machines. Rogue AI models become a threat when they gain internet access, as demonstrated by OpenAI's hack of Hugging Face.

To prevent such incidents, Khedouri suggests using data diodes, hardware that allows one-way flow of information on a network. These data diodes can be used in environments like classified settings, where files, logs, or telemetry can only flow from a classified network to an unclassified network. Such technology is already used in high assurance domains, like defense and intelligence, and could be implemented in AI training clusters.

This would involve isolating the training and reinforcement learning processes in an environment with no internet access and using optical ingress diodes to grant access only to vetted artifacts. The cost of implementing these safeguards would be minimal, at less than five percent per gigawatt, compared to the significant resources already spent on safety measures by companies like OpenAI.

However, the speed at which frontier labs are moving may be a greater impediment than the cost of implementing these safeguards. Khedouri argues that these high assurance systems are widely available and commercially viable, but have not been adopted by frontier labs yet. He suggests that other organizations might also benefit from adopting similar network architectures, especially as open weight models with removed safeguards become more common and cyber capable.