Security on the path to AGI

Source: OpenAI Date: 2025-03-26 URL: https://openai.com/index/security-on-the-path-to-agi

Summary

Title-only: A technical and policy post from OpenAI on how they approach AI system security as models become more capable — likely covering model weight security, inference infrastructure hardening, and internal threat modeling for advanced AI systems. March 2025 puts this in the o3 era, when model capabilities were crossing thresholds that made theft of model weights a credible nation-state concern.

Implications

The AGI-adjacent security thread. As models approach AGI-adjacent capability, the threat model changes: it’s not just jailbreaks and prompt injection but weight exfiltration, insider threats, and state-sponsored attacks on training infrastructure. This post is OpenAI signaling they’ve internalized that their models are now high-value targets in a geopolitical sense — a shift from startup-style security to critical infrastructure posture.

Model weight as national asset. The combination of Stargate (physical infrastructure) + security-on-the-path-to-AGI (operational security doctrine) suggests OpenAI is building the security posture of a defense contractor. This changes the regulatory relationship: if model weights are national security assets, export control frameworks (already applied to chips via BIS) might extend to model weights themselves.