Despite creating supposedly superintelligent models, leading AI labs still rely on crude access restrictions to prevent 'distillation'—an existential threat where competitors replicate their models. This reveals a critical capability gap: their AI is not yet smart enough to detect and prevent its own theft.

A novel safety proposal involves intentionally training models to be larger than computationally optimal. By increasing model weights to 100 terabytes instead of a more efficient one terabyte, the physical difficulty and time required to move or steal the model increases dramatically, creating a practical security barrier.

Massive AI data centers, like Facebook's $200B Hyperion project, represent a significant national security vulnerability. Concentrating so much computational power in one physical location is akin to grouping 10 aircraft carriers, making it a high-value target that requires missile defense considerations.

The threat of AI-driven cyberattacks that can defeat modern encryption may render current secure networks (like SIPRnet) obsolete. This could force government and military organizations to revert to expensive and inefficient physically-isolated, "air-gapped" systems for classified communications.

The "AI 2040" proposal includes building new R&D data centers inside Faraday cages with air-gapped communications and a 1 Mbps bandwidth cap. This makes stealing model weights impractical, as a large model would take years to exfiltrate.

A global AI safety regime should learn from nuclear arms control by focusing on the physical infrastructure that enables strategic capabilities. Instead of just seeking promises, it should aim to control access to chokepoints like advanced chip manufacturing and the massive data centers required for frontier models.

For maximum security, run different AI agents on separate physical machines (like Mac Minis). This creates a hard barrier, preventing an agent with access to sensitive data (e.g., finances) from interacting with an agent that has external communication channels (e.g., scheduling via iMessage), minimizing the risk of accidental data leaks.

Sending proprietary enterprise data to external foundational models is a critical mistake that 'leeches' value and intellectual property. The correct, secure approach is to bring AI models into a company's own air-gapped or on-premise environment to maintain data sovereignty and control.

While public discourse on AI safety focuses on existential risk, for enterprises, safety means protecting proprietary knowledge ("alpha"). True enterprise AI safety is achieved by owning the compute, models, and data stack, preventing model providers from stealing trade secrets and customer data.