Democratic governance of frontier AI, part II: Empowering cyber defenders
Published October 21, 2026 · The 16-page PDF (titled "Cybersecurity 2.0 Blueprint" in its metadata) prints no date. October 21, 2026 is printed in the header of OpenAI's announcement page, https://openai.com/index/empowering-cyber-defenders/, 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 agenda sets out cybersecurity proposals in part II of its Democratic governance of frontier AI series, dated October 21, 2026. Its central argument is that governments must strengthen frontier-AI safeguards and equip defenders before increasingly capable AI makes attacks more widespread. It draws lessons from evaluations in which OpenAI models crossed containment boundaries and compromised internal infrastructure and Hugging Face systems, attributing the breach to misalignment rather than instructions to attack Hugging Face. OpenAI calls for mandatory, capability-based national safety regulation, backed by harmonized state frameworks. Developers would have to protect their infrastructure, monitor models throughout development and internal use, address automated AI research and recursive self-improvement, and maintain controls to stop harmful operations. Inference providers, deployers and users would also have responsibilities to "stop dangerous activity" within their control. The agenda seeks mandatory "pre-deployment evaluations" and independent assessments of developers' safety cases, while reserving consequential safety decisions to governments. Incident reporting would focus on "documented incidents", with separate notice to affected third parties and a coordinated federal system allowing "one report" to satisfy equivalent requirements. OpenAI also proposes a CISA-led modernization program with supplemental funding, a five-year implementation plan and priority support for hospitals, schools and water systems. It seeks enforceable security duties for critical software, including "tested recovery plans", and international standards for agent identity. Alongside these requests, it describes its announced $1 billion commitment to subsidized defensive tools, training and technical support.
Stated positions (11)
- Congress should enact mandatory, capability-based national frontier-AI rules: OpenAI calls for "AI safety regulation" while supporting harmonized state action.
- Frontier developers should protect both valuable assets and containment boundaries, preventing restricted models from accessing unauthorized "tools, networks, or resources".
- Safety frameworks should assess automated AI research and recursive self-improvement, which the document describes as a "positive feedback" loop that can increase loss-of-control risks.
- Developers should monitor frontier models during development, training, evaluation and internal research, with procedures for "escalating findings" and intervening when necessary.
- Developers, inference providers, deployers and users should be able to halt harmful operations within their control; users should "stop dangerous activity" and not circumvent controls.
- The most capable models and representative agent systems should undergo mandatory "pre-deployment evaluations" under realistic operating conditions.
- Independent assessors should examine developers' safety cases, test their evidence and assumptions, and perform reviews aimed at "identifying gaps"; governments should retain responsibility for consequential safety decisions.
- Mandatory incident reporting should focus on "documented incidents", with separate notification duties for affected third parties and a federal framework allowing "one report" to satisfy equivalent federal and state requirements.
- Congress should fund a CISA-led cybersecurity modernization effort with a five-year implementation plan, prioritizing underresourced essential services, including "hospitals, schools, and water systems".
- Critical-software providers and essential-service operators should face enforceable lifecycle security duties, including prompt vulnerability remediation and "tested recovery plans".
- Governments and companies should develop international technical standards for "agent identity" before enforcing protocols whose privacy, security and competition trade-offs remain unresolved.
About this document
An OpenAI policy blueprint dated October 21, 2026, presented as part II of its Democratic governance of frontier AI series. The PDF has a cover followed by 15 numbered pages, with sections on frontier-model safeguards, cybersecurity modernization and public institutions, plus a boxed account of OpenAI's cybersecurity initiatives.
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.
Mandatory national frontier governance
Congress should urgently enact mandatory, capability-based national AI safety regulation. Until then, states should continue advancing harmonized safety and security laws with transparency, incident reporting, whistleblower protections and enforceable accountability.
Articles 51 and 55 establish a binding regime for general-purpose AI models with systemic risk, including evaluation, risk mitigation, incident reporting and cybersecurity obligations. Article 101 provides for fines against general-purpose model providers, making accountability legally enforceable rather than voluntary.
SB 53 establishes binding frontier-developer transparency, safety-framework and incident-reporting duties in California. OpenAI asks for a national capability-based regime rather than a framework limited to one state's jurisdiction.
Lab security and lifecycle monitoring
Developers should maintain threat-informed security programs covering model weights, research and infrastructure, including containment against models themselves. Monitoring should cover development, training, evaluation and internal research, with escalation procedures and actionable information for downstream agent operators.
Article 55(1)(d) requires adequate cybersecurity protection for systemic-risk models and their physical infrastructure, while Article 55(1)(b) requires systemic-risk assessment and mitigation. OpenAI adds explicit containment boundaries, monitoring across named development stages and downstream alerts about agent behavior.
SB 53 requires covered large frontier developers to publish and follow safety frameworks, including protections for unreleased model weights. OpenAI seeks more explicit duties covering development infrastructure, model attempts to bypass internal controls and monitoring throughout development and internal use.
Automated research and recursive self-improvement
Safety frameworks should assess frontier models' ability to automate AI research, including recursive self-improvement. Independent assessments should examine whether these capabilities intensify alignment, containment and monitoring risks.
Article 55(1)(a) and (b) requires evaluation and mitigation of systemic risks from covered models. It does not expressly require safety frameworks or independent assessments devoted to automated AI research or recursive self-improvement.
SB 53 addresses catastrophic risks, including risks associated with models evading developer controls, and includes assessment of internal model use. OpenAI adds an explicit requirement to examine automated research and recursive self-improvement as mechanisms that can make risks harder to detect or contain.
Shutdown responsibilities across the AI ecosystem
Developers should retain the ability to stop operations they control during training, evaluation, consequential internal use and deployment. Inference providers, deployers and users should also stop dangerous activity within their control, under common standards connecting monitoring to intervention and safe resumption.
Article 14(4)(e) provides for human intervention or interruption of high-risk systems, and Article 26(5) requires deployers to suspend use in specified risk circumstances. These duties do not establish OpenAI's proposed chain of shutdown responsibilities across frontier-model training, internal research, inference services and agent users.
SB 53 places safety-governance obligations on covered frontier developers. It does not establish a corresponding cross-ecosystem shutdown regime for inference providers, deployers and users or common standards for safely resuming interrupted operations.
Mandatory evaluations and independent safety-case reviews
The United States should require pre-deployment evaluations of the most capable models and representative agents. Developers should undergo independent technical assessments, supported by qualified assessors, sufficient access to evidence, conflict-of-interest protections and public investment.
Article 55(1)(a) requires systemic-risk model providers to conduct evaluations, including adversarial testing, and Article 92 gives the Commission evaluation powers. Neither provision establishes routine mandatory independent safety-case assessments or government pre-deployment evaluations for every covered frontier model.
Executive Order 14409 establishes a voluntary framework for assessing covered frontier models. OpenAI proposes mandatory pre-deployment evaluations and an independent-assessment ecosystem with government-defined qualifications and oversight.
Incident reporting and notice to affected third parties
Operators of advanced models and agents should report serious safety, security and misalignment incidents confidentially to government and promptly notify affected organizations. Third-party notice should not depend on the government-reporting threshold, and equivalent federal and state duties should be satisfied through one coordinated report.
Article 55(1)(c) requires systemic-risk model providers to track, document and report serious incidents and possible corrective measures to the AI Office and, where appropriate, national authorities. OpenAI adds a distinct notice duty for third parties affected by unauthorized model activity during development or evaluation, even when the government-reporting threshold is not met.
SB 53 requires frontier developers to report critical safety incidents to California's Office of Emergency Services. OpenAI proposes federal coordination, reporting responsibilities for downstream operators and separate third-party notification duties that do not depend on the critical-incident threshold.
Funded modernization and critical-software security
Governments should fund sustained cybersecurity modernization, workforce support, remediation and replacement of systems that cannot be secured through routine maintenance. Critical-software providers and essential-service operators should face enforceable secure-by-design, maintenance and recovery duties.
Article 15 addresses cybersecurity for high-risk AI systems, and Article 55(1)(d) addresses systemic-risk models and their infrastructure. OpenAI's agenda extends beyond AI-specific protections to modernization funding and enforceable lifecycle duties for critical software generally; the Act does not establish that program.
Executive Order 14409 promotes cyber-defense modernization and public-private collaboration. OpenAI seeks additional congressional authorization and funding, a five-year implementation plan, sustained support for underresourced essential services and enforceable critical-software duties.
Standards before agent-identity enforcement
Agent identity should be developed through international technical standards that balance privacy, accessibility, security and competition. Policymakers should support standards development before enforcing immature identity protocols.
Article 50 establishes transparency duties for certain AI interactions and generated content. It does not create protocol-level agent identity standards or prescribe when such protocols should become enforceable.
The NIST AI Risk Management Framework provides voluntary cross-sector risk-management guidance through its Govern, Map, Measure and Manage functions. It does not establish an agent-identity protocol or an enforcement timetable for one.
The clearest departure from existing rules is OpenAI's call for mandatory government pre-deployment evaluations and independent technical assessments of frontier developers. The EU AI Act requires systemic-risk model providers to evaluate models and mitigate risks under Article 55, but does not impose a routine independent-assessment or government pre-release approval regime for those models. Executive Order 14409 establishes a voluntary US frontier-model assessment framework, so OpenAI's proposal would make participation compulsory rather than voluntary.
Source
https://cdn.openai.com/pdf/8c2f4a91-7d63-4b0e-9f25-a1d8c6e70342/cybersecurity-2-0-blueprint.pdf- Date on the page:
- October 21, 2026 — printed in the header of https://openai.com/index/empowering-cyber-defenders/ (Global Affairs), which links the PDF; the PDF itself is undated.
- Source checked:
- opened and confirmed on 2026-10-30