Frontier Governance Framework
Published May 28, 2026 · The 22-page PDF prints no date. May 28, 2026 is printed on OpenAI's announcement page, https://openai.com/index/openai-frontier-governance-framework/, which links this PDF.
Not law. This is a company's own public position on AI regulation. It is not law, and it carries no legal force.
What it argues for
OpenAI's Frontier Governance Framework, dated May 28, 2026, documents the company's technical and organizational processes for managing risks from frontier models. It is a framework for OpenAI's own conduct, designed to meet "baseline legal requirements" under California's Transparency in Frontier AI Act and to serve as a public summary of its Safety & Security Framework under the EU General-Purpose AI Code of Practice. It covers cyber offense, chemical, biological, radiological and nuclear threats, harmful manipulation, and loss of control. Risk assessments combine capability evaluations, threat modelling, expert input and post-deployment monitoring. Decisions incorporate safety margins and a holistic judgment about residual risk rather than relying solely on evaluation scores. Models with excessive residual risk are "not deployed unless" further mitigations sufficiently reduce that risk. The approach is less developed for harmful manipulation, which "remains exploratory" and may be addressed primarily through system-level safeguards and monitoring. The framework also describes incident investigation and legally required reporting, layered security controls protecting model weights and other assets, and publication of assessment details in system cards or other launch reporting. External expert involvement is discretionary. OpenAI assigns US compliance responsibility to OpenAI OpCo LLC and EU provider responsibility to OpenAI Ireland Limited, whose board oversees systemic risk for EU purposes. Framework assessments occur "at least once every 12 months", while material changes and their justifications must be published "within 30 days of the update".
Stated positions (11)
- The framework establishes a legal-compliance baseline for frontier-model governance: it is designed to meet "baseline legal requirements" and document "technical and organizational processes".
- Risk identification covers severe harm across model development, storage, use and deployment, including incidents involving "greater than 50 fatalities" or "$1 billion of" property damage or losses.
- Deployment decisions consider uncertainty and safety margins: evaluation findings contribute to "a holistic judgment" about whether a capability threshold has been reached.
- Models must not be deployed with excessive residual risk without further safeguards: the model is "not deployed unless" additional mitigation sufficiently reduces risk.
- Harmful-manipulation assessment is unfinished: the approach "remains exploratory", and OpenAI favours measures including "post-deployment monitoring".
- Incident handling combines detection, investigation, containment and external reporting: reportable incidents go to appropriate authorities "within the required deadlines".
- Security protects model weights, training data and customer data through layered controls: "Access is tightly managed" and "environments are continuously monitored".
- OpenAI publishes assessment details at launch: reporting appears in "system cards or in other" reporting, alongside internal Safety and Security Model Reports.
- External evaluation is discretionary rather than universal: OpenAI "may solicit and obtain input" from external experts, including "independent third-party evaluators".
- The framework assigns compliance responsibility to separate US and EU entities and states that OpenAI Ireland Limited's directors "exercise systemic risk oversight" for EU purposes.
- The framework receives periodic review and public change documentation: assessments occur "at least once every 12 months", and material updates are published "within 30 days of the update".
About this document
OpenAI's Frontier Governance Framework is a PDF dated May 28, 2026, with a cover, contents page and 20 numbered pages. Seven main sections cover systemic risk assessment and mitigation, security, model reporting, external experts, allocation of responsibility and framework change management.
How this sits against AI law
Each stance compared with what EU and US instruments actually require. Where no instrument addresses a theme, that gap is shown rather than hidden.
Lifecycle systemic-risk assessment
OpenAI assesses severe risks using threat modelling, capability evaluations, external information and post-deployment monitoring. Its principal categories are cyber offense, CBRN, harmful manipulation and loss of control, although harmful-manipulation assessment remains exploratory.
Article 55(1)(a) and (b) requires providers of general-purpose AI models with systemic risk to conduct model evaluations, including adversarial testing, and assess and mitigate systemic risks. OpenAI's lifecycle assessment processes address these duties; describing an exploratory approach to one category does not establish that the duty has been fully met.
SB 53's frontier AI framework provisions require large frontier developers to describe their protocols for assessing and mitigating catastrophic risks. OpenAI presents this framework as its implementation of that obligation, rather than as a request for new legislation.
Deployment gate for unacceptable residual risk
OpenAI will not deploy a model whose residual risks exceed acceptable levels unless additional mitigation sufficiently reduces the risk. Acceptance decisions incorporate safety margins, evaluation results and other evidence.
Article 55(1)(b) requires assessment and mitigation of systemic risks, but does not prescribe OpenAI's risk tiers, acceptance thresholds or this express deployment gate. The framework adds a specific internal decision rule to the statutory mitigation duty.
SB 53 requires publication of a frontier AI framework and compliance with its terms, but does not itself establish a universal acceptable-risk threshold or an equivalent categorical deployment prohibition. OpenAI's gate is an additional substantive commitment, even though the law can make adherence to the published framework consequential.
Safety incident response and reporting
OpenAI maintains incident-response plans, investigates potential incidents and determines whether external reporting is legally required. Reportable incidents are sent to the appropriate authorities within applicable deadlines.
Article 55(1)(c) requires providers of systemic-risk models to track, document and report relevant information about serious incidents and possible corrective measures to the AI Office and, where appropriate, national competent authorities without undue delay. The framework commits to applicable reporting deadlines without specifying a separate EU timetable.
SB 53 requires reporting of critical safety incidents to California's Office of Emergency Services, generally within 15 days and within 24 hours for incidents presenting an imminent risk of death or serious physical injury. OpenAI's commitment to report within legally required deadlines matches that obligation.
Protection of model weights and infrastructure
OpenAI applies layered security controls to model weights, training data and customer data. Measures include encryption, access restrictions, monitoring, personnel screening, sandboxing and internal and external security assessments.
Article 55(1)(d) requires an adequate level of cybersecurity protection for systemic-risk models and their physical infrastructure. The described controls implement that objective, although the Act does not individually mandate every security technique named in the framework.
SB 53's frontier AI framework provisions address technical and organizational protocols for catastrophic-risk management, including security measures. OpenAI documents concrete controls within that framework; the statute does not prescribe each listed encryption, screening or testing method.
Public model-risk reporting
OpenAI documents assessment and mitigation results in Safety and Security Model Reports and publishes assessment details through system cards or other launch reporting. It also reviews whether reports for its most capable frontier models need updating every six months, subject to stated exceptions.
Articles 53(1)(a) and 55 require technical documentation and systemic-risk management, while Article 53(1)(d) separately requires a public training-content summary. They do not generally require publication of systemic-risk assessment details in a launch system card, so that public-facing commitment goes beyond the Act's express disclosure duties.
SB 53 requires large frontier developers to publish transparency reports concerning catastrophic-risk assessments and mitigation for covered models. The framework explicitly identifies its model reports with that reporting regime and commits to publishing assessment details.
Governance and framework updates
OpenAI assigns US and EU compliance responsibility to named corporate entities and gives OpenAI Ireland Limited's board EU systemic-risk oversight. It commits to annual framework assessments, board oversight of material changes and publication of changes and justifications within 30 days.
Article 55 establishes provider responsibility for systemic-risk management but does not prescribe annual assessments of a public governance framework or publication of a material-change log within 30 days. These are more specific governance commitments than the Act itself sets out; the voluntary Code of Practice supplies a separate compliance route under Articles 55(2) and 56.
SB 53 requires covered large developers to maintain and update a public frontier AI framework and disclose material modifications within 30 days. OpenAI's periodic assessment and public-change process implements those framework-maintenance obligations.
The clearest distinction is between transparency obligations and a substantive deployment gate. California's Transparency in Frontier AI Act requires covered developers to publish and follow a frontier AI framework, but does not itself prescribe a universal acceptable-risk threshold or prohibit every deployment above one. OpenAI adds its own commitment that a model with excessive residual risk is "not deployed unless" additional mitigations sufficiently reduce the risk. The EU AI Act's Article 55 requires systemic-risk assessment and mitigation, but does not express that duty through OpenAI's particular risk-tier and acceptance process.
Source
https://cdn.openai.com/pdf/e37d949b-8c9f-4d76-b99e-4272f4631a7e/openai-frontier-governance-framework.pdf- Date on the page:
- May 28, 2026 — printed in the header of https://openai.com/index/openai-frontier-governance-framework/ (category Safety), which links the PDF; the PDF itself is undated.
- Source checked:
- opened and confirmed on 2026-10-30