OpenAI wrote the exam it wants to sit

OpenAI has published the first document that maps a frontier lab's internal safety processes directly onto California's SB 53 and the EU AI Act's Code of Practice for General-Purpose AI. The interesting thing about the Frontier Governance Framework is not what it commits OpenAI to. It is what it commits everyone else to explaining.

The framework landed on 28 May. It restates OpenAI's existing Preparedness Framework (the internal risk-scoring process the company introduced in 2023) and translates it into four risk domains — cyber offence, CBRN (chemical, biological, radiological, nuclear), harmful manipulation, loss of control — and four procedural pillars: model reporting, security risk management, incident response, and external expert input.¹ No other frontier lab has published an equivalent dual-mapping document. Anthropic, Google DeepMind and xAI now have to either match the structure or explain why they have not.

The first-mover move. This is standard-setting dressed as compliance. Regulators and auditors reach for whatever reference document exists when they need to decide what "adequate" looks like. SB 53 (the Transparency in Frontier Artificial Intelligence Act) requires large frontier developers to publish a framework explaining how they identify and govern catastrophic risks; it does not specify the shape of that framework.² OpenAI has just specified it. The shape now exists. Deviation becomes the thing that needs defending.

The Illinois timing matters. Illinois SB 315, mandating third-party AI audits, passed the same week.² An audit regime needs a checklist. The checklist does not exist yet. OpenAI has now offered one — not formally, not in those words, but functionally. If an auditor in 2027 needs to assess a frontier model against SB 315, the document on the desk will be the one that already maps risk domains to procedural pillars in regulator-legible language. That document is OpenAI's.

This is defensible as good-faith coordination with regulatory intent. Companies with operational knowledge of how models actually fail are well placed to describe what monitoring should look like. The same fact pattern is defensible as the canonical regulatory capture move: the regulated entity drafts the rules of its own examination, then submits a perfect answer. Both readings are supported by the evidence. The honest position is that they are the same action viewed from different angles, and which angle matters depends on what the framework actually constrains.

What it does not constrain is the interesting part. Anthropic's Responsible Scaling Policy (RSP) ties specific commitments to capability thresholds. A model reaching ASL-3 triggers mandatory security measures and implicitly halts further scaling until conditions are met.³ The commitment is irreversible in the sense that crossing the threshold without the controls in place would be a public violation of a published policy.

OpenAI's framework covers similar risk categories without similar trip-wires.

OpenAI's framework covers similar risk categories without similar trip-wires. It is organised around reporting, management processes, and external expert input. There is no published commitment that a specific capability evaluation result triggers a specific operational consequence. "External expert input", and this is true across most frontier lab safety documents, means advisory access, not veto power.¹ The governance question regulators and civil society care about most, which is who has authority to halt a deployment, is not directly answered.

The process-versus-threshold defence is real, though. It would be unfair not to make it. Threshold-based commitments depend on the accuracy of the capability evaluations triggering them, and capability evaluations are a contested research area. Anthropic's RSP has faced legitimate criticism that its threshold definitions are too loosely specified to be externally verifiable.³ A process-based framework, published reporting, incident response, structured external input, is more auditable in a conventional sense. You can check whether reports were filed. You cannot easily check whether a capability threshold was correctly assessed.

Process compliance, in other words, is easier to grade. Whether it is what we actually want graded is a different question.

The lobbying loop is published, not hidden. OpenAI engaged publicly with both SB 53 and the EU GPAI Code during their drafting phases. The framework now maps to those same instruments.² This is not a scandal; it is a closed circuit. The company helped shape the rules, then published a document demonstrating its alignment with them. Whether that is efficient policy coordination or self-dealing depends on whether the rules, as shaped, actually bind. The framework's structural lightness on irreversible commitments suggests one answer. The framework's genuine procedural seriousness on incident response and reporting suggests another. Both are present in the same document.

The scope question. Both regulatory instruments the framework maps to are US and EU. SB 53 is Californian. The EU's GPAI Code of Practice is a soft-law instrument under the EU AI Act — providers sign it to demonstrate compliance, and it shapes enforcement calibration rather than constituting binding law in the same way as direct regulation.² A framework that covers SB 53 and the GPAI Code is a complete answer to the regulatory question as posed in democratic markets with active AI legislation. It says nothing about deployment in jurisdictions without equivalent regimes. The framework does not claim to fill that gap, and it would be unreasonable to demand it does. But the gap is the part of the map where the most consequential deployments may end up taking place, and naming it matters.

What to watch. Three things. First, whether Anthropic, Google DeepMind and xAI publish their own dual-mapping documents within the next quarter, and whether they adopt OpenAI's four-domain / four-pillar structure or push back on it with an alternative. Second, whether the Illinois audit regime, when it produces implementation guidance, cites or echoes the framework's vocabulary. Third — and this is the one that determines whether the document is governance or marketing — whether any future OpenAI model evaluation produces an outcome under the framework that visibly changes a deployment decision. If the framework has been operating for twelve months and no deployment has been altered by it, the structural lightness was the point.

The Frontier Governance Framework is well-written, regulator-legible, and probably more useful than the equivalent silence from rivals. It is also a document in which a lab that helped shape the rules has published the first authoritative answer to the test those rules pose. Both things are true, and pretending only one of them is true is the analytical mistake.

Footnotes

¶