OpenAI Is Starting to Look Like a Government

Creative Robotics
OpenAI Is Starting to Look Like a Government

Scroll through OpenAI's announcements from just the past couple of weeks and a strange pattern emerges. There's the appointment of Jill Hazelbaker as Chief Corporate Affairs Officer. There's Dan Driscoll coming aboard as Senior Counselor for National Security Modernization. There's a new National Community Compact and Compute Campus Policy Blueprint governing how data centers interact with local communities. There's the Grid Protectors Program, designed to work with utilities so AI compute campuses don't destabilize electricity costs for ordinary consumers. There's an expanded bug bounty program with loyalty bonuses for security researchers, a hardening sprint with Cathedral Systems, and a specialized cybersecurity model, GPT-6 Cyber, built explicitly to patrol the boundaries of authorized scope.

None of these moves, taken individually, is shocking. Companies hire comms executives. Infrastructure projects need community relations. Security researchers deserve better pay. But stacked together, they describe something unusual: a private AI lab assembling the institutional scaffolding of a state actor. National security counselors are the kind of hires you expect from defense contractors or federal agencies, not software companies. Grid reliability programs are the kind of commitments utilities extract from heavy industrial users, not typical cloud tenants. A 'Community Compact' is language borrowed from municipal governance, not product marketing.

This matters because OpenAI's own research is simultaneously demonstrating just how much structural power these systems already wield. The company's pipeline that autonomously discovered and patched 948 Linux kernel privilege-escalation vulnerabilities—detailed in 'A Thousand Fewer Paths to Root'—is a genuine security achievement, but it's also proof that an AI lab can now unilaterally probe and alter the foundations of global computing infrastructure at a scale no single government agency can match. Pair that with GPT-6 Cyber's dual-use nature—finding exploits and fixing them—and you have a company that functions less like a vendor and more like a sovereign cybersecurity authority, accountable mostly to itself.

The updated Model Spec, with its language about 'human agency' and 'faithfulness to human intent,' reads almost like a constitutional document for a jurisdiction that doesn't officially exist. Meanwhile, the fair use court brief campaign and the new robots.txt signal for AI training preferences show OpenAI positioning itself as an interlocutor with creators and publishers on terms it helped define.

None of this is necessarily sinister. Compute campuses really do strain power grids. Security researchers really are underpaid. National governments really do need advisors who understand frontier AI's security implications. But there's a difference between a company responding to its growing footprint and a company quietly building permanent institutional machinery—comms, diplomacy, grid policy, national security liaison roles—that normally belongs to public entities with electoral accountability.

As AI labs take on more state-like functions, the question isn't whether they're doing it competently. It's who gets to decide they should be doing it at all, and what checks exist when a corporation's security sprint affects infrastructure the rest of us didn't choose to depend on.