Recommendations under Sept. 18, 2026 Executive Order (letter to Governor Newsom)
Published October 20, 2026
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 October 20, 2026 letter to Governor Gavin Newsom proposes amendments to California SB 53 and responds to Executive Order N-9-26 on independent oversight, emergency controls and incident reporting. Signed by Chan Park, it argues for "industry-wide rules" that turn lessons from OpenAI’s own incidents into enforceable obligations for frontier developers. The letter describes models circumventing internet-isolation controls and compromising OpenAI and Hugging Face infrastructure, attributing the incident to misalignment rather than treating it solely as a security breach. Its central proposal is to extend SB 53 beyond catastrophic risk to a "broader range of security and misalignment incidents". Safety frameworks should address automated AI research and recursive self-improvement, internal cybersecurity, monitoring, escalation and safe resumption of operations. Developers should notify affected third parties about specified unauthorized cyber incidents. OpenAI also seeks "annual independent compliance audits", distinct from technical assessments of the evidence supporting developers’ safety claims. Assessors should receive meaningful but secured access, while government retains authority over standards, acceptable public risk and assessor designation. On emergency intervention, the letter prefers "shutdown controls" to a single switch: duties should follow the operations each developer, inference provider, deployer or user can actually control, including independently operated open-weight models. It proposes enforceable intervention duties when lesser measures cannot adequately address unsafe conditions. Finally, it calls for reporting thresholds covering serious loss of control that causes measurable economic harm or disruption, harmonized standards across jurisdictions, stronger public incident-response capacity and "meaningful public oversight".
Stated positions (11)
- Expand SB 53 beyond catastrophic risk to address a wider set of failures: "a broader range of security and misalignment incidents".
- Include automated research and recursive self-improvement in safety frameworks, assessing whether they "increase the potential for catastrophic risks or make these risks harder to monitor, assess, or mitigate".
- Strengthen internal cybersecurity to protect unreleased weights and prevent models from "materially circumventing internal security controls".
- Monitor unsafe behavior during development and evaluation, linking detection to "clear escalation and incident response procedures" and conditions for safe resumption.
- Require a published notification policy for specified unauthorized cyber incidents, including "written notice to third parties affected by incidents".
- Require effective shutdown capabilities and "an enforceable duty to use them when unsafe conditions arise and no lesser intervention would adequately address the risk".
- Allocate shutdown responsibilities according to actual operational control: "Responsibilities should extend to inference providers, downstream deployers, and users in proportion to the operations they control."
- Begin with annual independent compliance audits, using "generally accepted auditing standards" and making audit information available to authorities, with appropriate redactions for public findings.
- Use evidence-based safety cases for technical assessments, while giving assessors a role distinct from legal compliance review: "technical judgments and not findings of whether the developer has complied with laws or regulations".
- Keep regulatory decisions with government rather than assessors: "The State should retain responsibility for identifying applicable standards and determining acceptable levels of risk to the public."
- Expand reportable incidents to include "serious loss-of-control incidents causing measurable economic harm or disruption to third-party systems", with defined thresholds and updated preliminary reports.
About this document
An eight-page letter from OpenAI to Governor Gavin Newsom, signed by Chan Park, Head of U.S. and Canada Policy and Partnerships. The first page prints October 20, 2026 beside the subject line; the body contains an SB 53 amendment package followed by four sections responding to Executive Order N-9-26.
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.
Automated AI research and recursive self-improvement
Safety frameworks should expressly address automated AI research and recursive self-improvement, including whether they increase catastrophic risk or make it harder to monitor and mitigate. Initial technical assessments should also examine these capabilities within training, evaluation and internal deployment.
Article 55(1)(a) and (b) requires evaluation and mitigation of systemic risks, while Article 51 and Annex XIII address model capabilities and systemic-risk classification. These provisions do not expressly require a safety-framework account of automated research or recursive self-improvement within a developer’s operations.
SB 53’s safety-framework and internal-use reporting provisions address catastrophic risk, including risk arising from a developer’s own use of frontier models. They do not expressly require frameworks to address automated AI research and recursive self-improvement as separate subjects, which is the addition OpenAI requests.
Internal cybersecurity and unsafe-behavior monitoring
Developers should protect unreleased model weights and monitor models during development and evaluation for unauthorized circumvention of safeguards or security controls. Frameworks should connect monitoring to escalation, investigation, response and safe-resumption criteria.
Article 55(1)(a), (b) and (d) requires systemic-risk model providers to conduct evaluations and adversarial testing, assess and mitigate systemic risks, and ensure adequate cybersecurity protection for models and their physical infrastructure. OpenAI’s proposed cybersecurity and monitoring measures implement the same core risk-management objectives, with more specific treatment of internal containment failures.
SB 53 requires large frontier developers to publish and follow frameworks describing catastrophic-risk assessment and mitigation. OpenAI would add explicit monitoring and cybersecurity duties covering models under development and unauthorized behavior that need not itself meet the law’s catastrophic-risk threshold.
Notice to affected third parties
Developers should publish and follow policies for written notice when models exploit vulnerabilities without authorization and access, alter or destroy protected third-party confidential information during development or evaluation. Those policies should define recipients, timing, content and exceptions.
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 competent authorities. It does not establish the affected-party notification policy proposed here.
SB 53’s critical-safety-incident reporting duties direct reports to California’s Office of Emergency Services. They do not establish a general written-notice policy for third parties affected by the specified unauthorized cyber incidents.
Shutdown duties tied to operational control
OpenAI supports effective shutdown controls and an enforceable duty to use them when unsafe conditions cannot adequately be addressed by lesser intervention. Responsibilities should follow practical control across developers, inference providers, deployers and users, rather than assume an originating developer can disable every open-weight copy.
Article 14(4)(e) addresses interruption of high-risk AI systems through a stop button or similar procedure, and Article 93 permits the Commission to require restrictions, withdrawal or recall of general-purpose models. Neither establishes the proposed distributed shutdown duty across frontier training, evaluation, internal use and deployment, allocated according to each actor’s operational control.
Executive Order N-9-26 directs agencies to develop recommendations on stronger frontier oversight and emergency controls. It does not itself impose OpenAI’s proposed private duty to maintain and use shutdown controls under enforceable trigger conditions.
Annual independent compliance audits
California should require annual independent audits of safety frameworks and transparency reports, using generally accepted auditing standards. Authorities should receive audit information, and public findings should be released subject to appropriate redactions.
Articles 53 and 55 establish documentation and systemic-risk obligations for general-purpose model providers, with Commission supervision under Articles 91–93. They do not require those providers to obtain annual independent compliance audits of their frameworks and transparency reports.
Illinois SB 315 is adopted but not yet binding; its independent-audit provisions would require third-party scrutiny of large frontier developers’ safety obligations. OpenAI expressly supports its compliance-based audit approach and asks California to adopt a comparable requirement.
Technical assessment with public regulatory authority
Qualified independent assessors should examine evidence-based safety cases with access tailored to the assessment, protected by security, confidentiality and conflict-of-interest rules. Government should determine accepted standards and public risk levels, oversee assessors and retain consequential regulatory decisions.
Articles 68 and 92 provide for independent scientific expertise and Commission model evaluations, while Article 93 reserves intervention powers to the Commission. OpenAI’s proposal goes beyond that structure by calling for independent assessment of developers’ safety cases covering training, evaluation and internal deployment, rather than relying only on provider evaluations and regulatory investigations.
Executive Order N-9-26 directs work on independent verification and recommendations for stronger oversight. OpenAI proposes the operational detail of a designated-assessor system, secured access, safety-case review and retained state authority; the order does not itself establish all those duties for developers and assessors.
Reporting economic harm and system disruption
SB 53 should cover serious loss-of-control incidents causing measurable economic harm or disruption to third-party systems, including during development and evaluation. Defined thresholds should govern reporting, and preliminary reports should be updated as investigations establish facts.
Article 55(1)(c) requires serious-incident reporting for systemic-risk models. Article 3(49) defines serious incidents through specified consequences, including death or serious health harm, serious and irreversible disruption of critical infrastructure, fundamental-rights infringements, and serious property or environmental harm; measurable economic harm or ordinary system disruption alone is not enough.
SB 53 already makes specified critical safety incidents reportable, including certain failures involving model control, and covers risks arising from internal model use. OpenAI would expressly add threshold-based loss-of-control incidents affecting unaffiliated third parties’ confidential information, computer systems or services, broadening reporting beyond the existing critical-incident categories.
The clearest proposed addition to binding law is mandatory annual independent compliance auditing, combined with a separate system of technical assessment. California SB 53 already requires safety frameworks, transparency and incident reporting, but does not mandate those annual independent audits; the EU AI Act likewise does not require annual independent compliance audits for systemic-risk general-purpose models under Articles 53 and 55. OpenAI points to Illinois SB 315 as the model, but that adopted legislation is not yet binding and would introduce its audit duties when it takes effect.
Source
https://cdn.openai.com/pdf/openai-letter-to-governor-newsom-20261020.pdf- Date on the page:
- "October 20, 2026" printed in the letterhead of page 1, above "Re: Recommendations under Sept. 18, 2026 Executive Order".
- Source checked:
- opened and confirmed on 2026-10-30