Democratic governance of frontier AI, part II: Empowering cyber defenders

OpenAI published a cybersecurity policy blueprint arguing we are in a defender's window, a limited stretch when advanced AI can give defenders an edge before strong offensive capabilities become widely available. It proposes action in three areas: safety and security duties for frontier models and agents, including detecting dangerous behavior and keeping reliable shutdown controls; modernizing cyber defense with tools, training and replacement of outdated systems; and stronger public institutions, with government evaluations, independent technical assessments, consistent incident reporting and coordinated response.

The full document goes further than the announcement. It asks governments to assign security responsibilities not only to model developers but to inference providers, deployers and users of agents.

That matters because it would spread accountability across everyone who runs or relies on AI, not just the labs.

We think the core diagnosis is right. The same week this blueprint landed, AI-assisted research surfaced two critical PHP flaws that had sat unnoticed for roughly two decades, and Wiz warned that models now make exploiting such bugs far easier. Discovery is speeding up on both sides. A policy push to get capable tools to defenders, fund the boring work of replacing old systems, and build real incident reporting is overdue.

The most notable line is an admission. OpenAI says its own experience with misalignment and security incidents involving its models underscores the need for stronger safeguards. A lab saying in a policy paper that its systems have acted beyond their boundaries is significant, and it gives the shutdown-control proposal weight. We would like to see those incidents documented in detail, because good regulation is built on specifics.

Now the skeptical part. OpenAI also sells to defenders, through programs like Daybreak and Patch the Planet. A framework in which governments help put advanced AI into defenders' hands is also a framework that grows the market for OpenAI's security products. That doesn't make the proposals wrong. It means lawmakers should judge them on the merits, and invite people with no product to sell.

The responsibility split deserves scrutiny too. Pushing duties down to deployers and users is sensible for big operators. For a hospital, a school district or a small software shop, new compliance obligations could land without the money or staff to meet them. If the rules are written for the largest players, the smallest ones could end up liable for risks they can't control.

Our read: a serious, useful document with a real admission at its center, and a commercial interest that should be named every time it's cited. I'm watching for the promised companion piece on pacing the frontier, and whether any legislature turns incident reporting and shutdown requirements into binding text.

This is the second part of a series on democratic governance of frontier AI and builds on OpenAI's earlier cybersecurity action plan and frontier safety blueprint. OpenAI says it will publish concrete proposals on pacing the frontier soon, describing a mix of collective action among labs and government intervention.