Meta Superintelligence Scaling Framework — Version 2.1
Published October 2026 · The PDF prints only a month: "Version 2.1 | October 2026" in its running header and a "Preface to Version 2.1 (October 2026)". Meta's announcement post, https://research.meta.ai/blog/developing-capable-models-responsibly, is dated October 2, 2026; that day is used for sorting only.
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
Meta’s Superintelligence Scaling Framework, Version 2.1, sets the company’s internal rules for managing catastrophic risks throughout frontier-model development and release. Printed October 2026, it revises and renames the Advanced AI Scaling Framework published in April 2026. Its central approach is to assess whether a model could "substantially contribute" to threat scenarios in cybersecurity, chemical and biological weapons, and loss of control, then validate safeguards before proceeding. High-risk deployments and critical-risk development may continue only with safeguards that reduce risk to the "moderate or lower risk threshold". Version 2.1 adds a third loss-of-control outcome covering containment during training and evaluation. It calls for prospective capability assessments, approved and red-teamed sandboxes, vulnerability remediation, and "Real-time, write-only chain-of-thought and trajectory logging". High-risk training and evaluation require monitoring of all rollouts, subject to senior approval of exceptions; automated systems capable of stopping individual rollouts remain a development commitment. The revision also expands open-weight release assessments: Meta recognizes research and transparency benefits but acknowledges that adversaries can "fine-tune refusal behavior away", so assessments must consider continued training and realistic deployment contexts. Existing commitments include release-linked preparedness reports, internal-use risk assessments, a published model specification, and named executive responsibility. The framework says more formal board governance is underway; the accompanying announcement promises a "new AI committee of our Board of Directors" to review framework changes and check operational conformity. Meta connects the update and planned governance changes to the voluntary White House Accord on Super Intelligence, rather than presenting them as statutory requirements.
Stated positions (12)
- Risk management covers development as well as release: the framework applies "from training to evaluation to deployment", including internal, closed and open-weight settings.
- Catastrophic-risk thresholds depend on a model’s contribution to threat scenarios, not capability alone: Meta tests whether it could "substantially contribute" to those scenarios.
- Critical-risk development is conditional, not automatically prohibited: Meta will continue only when assessments are complete and safeguards are validated to "reduce risk to the moderate or lower risk threshold".
- Containment risk must be assessed before training or evaluation: "The foreseeable high severity risks identified through this assessment then must be mitigated prior to training or evaluation."
- Sandbox vulnerabilities must be addressed before specified risky training runs: Meta will "require remediation of all detected medium to high severity vulnerabilities before initiating new high- or critical-risk training runs with the impacted sandboxes".
- Training and evaluation records must be protected against alteration: "All model rollouts must be logged to write-only data stores in real time".
- High-risk runs require monitoring, but exceptions remain possible: they require "chain-of-thought and trajectory monitoring to be active for all rollouts", with exceptions approved by the Director of Alignment and Risk or Chief AI Officer.
- Open-weight assessments must account for safeguards being removed: Meta acknowledges that adversaries can "resample to bypass refusals, prefill outputs, or fine-tune refusal behavior away".
- Release decisions must be explained publicly: preparedness reports will describe "risk assessment, evaluation results, implemented mitigations, and rationale for deployment decisions for each risk domain".
- Board-level AI oversight is planned rather than already established: the announcement says Meta "will establish a new AI committee of our Board of Directors".
- The framework’s catastrophic-risk scope excludes some serious harms: risks must satisfy four criteria, including "Net new" and "Instantaneous or irremediable"; other harms are addressed through separate processes.
- The framework must be reassessed and updated regularly: Meta commits to updating it when adequacy or adherence is materially undermined, or "at least every twelve months, whichever is sooner".
About this document
A 51-page Meta PDF headed Meta Superintelligence Scaling Framework and marked Version 2.1 | October 2026. It contains a preface, five main sections, a terminology appendix and a change log; the accompanying Meta AI Research announcement explains the revision and planned board AI committee.
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.
Containment and monitoring before release
Version 2.1 adds prospective containment assessments before training or evaluation, sandbox validation, protected rollout logging and monitoring for high-risk runs. Monitoring exceptions need senior approval, and automated rollout shutoff systems are still to be developed.
Article 55(1)(a), (b) and (d) requires evaluation, systemic-risk assessment and mitigation, and adequate cybersecurity for systemic-risk models. It does not prescribe Meta’s particular training-time sandbox approval, write-only logging or rollout-monitoring architecture.
SB 53’s frontier-framework provisions require covered large developers to describe catastrophic-risk assessment and mitigation practices, including relevant security safeguards. The statute does not prescribe these particular sandbox, logging and monitoring controls for training and evaluation.
Open-weight release risks
Meta supports open-weight releases for research and transparency but assesses their distinct misuse risks. Evaluations consider adversaries who can alter model behavior through continued training, rather than relying only on refusal behavior in the released model.
Article 53(2) excludes general-purpose AI models with systemic risk from the free and open-source exemption to specified documentation duties. Article 55 therefore preserves systemic-risk evaluation and mitigation duties regardless of an open release; it does not prescribe Meta’s precise adversary-testing methods.
Pillar I’s section on encouraging open-source and open-weight AI supports their research and innovation benefits while treating release as a developer decision. The Action Plan is a policy roadmap, not a binding obligation on Meta to conduct open-weight safety assessments.
Conditional development and deployment
High-risk models may be deployed, and critical-risk models may continue development, only after sufficient safeguards are defined, implemented and validated. The detailed critical-risk rule requires complete assessments and reduction to moderate or lower risk.
Article 55 requires providers of systemic-risk models to evaluate models and assess and mitigate systemic risks; it does not automatically prohibit development or deployment because a model presents systemic risk. Meta’s conditional approach matches that mitigation-based structure, although its internal risk categories are not legal compliance thresholds.
SB 53’s frontier-framework provisions address catastrophic-risk assessment, mitigation and deployment decision-making. They do not establish an automatic prohibition corresponding to Meta’s high or critical categories, or make Meta’s moderate-or-lower rating a statutory safe harbor.
Limits on catastrophic-risk scope
The framework concentrates on cybersecurity, chemical and biological risks, and loss of control. Its general inclusion criteria require plausible, catastrophic, net-new and instantaneous or irremediable outcomes; other serious harms are assigned to separate company processes.
Article 3(65) defines systemic risk to include large-scale adverse effects on public health, safety, public security, fundamental rights or society as a whole. Article 55(1)(b) requires assessment and mitigation of systemic risks without limiting them to net-new or instantaneous or irremediable outcomes, so this framework alone covers less than the legal duty.
SB 53’s catastrophic-risk definition focuses on specified severe consequences arising from frontier capabilities, including weapons assistance, cyberattacks and evasion of developer control. It does not make net-new harm or instantaneous or irremediable consequences prerequisites, so Meta’s inclusion criteria are narrower than the statute’s risk scope.
Planned board AI committee
Meta already assigns evaluation and mitigation responsibility to its Chief AI Officer and Director of Alignment and Risk. Its announcement adds a planned board AI committee to review future framework changes and independently check that operations conform to company standards.
Article 55 places systemic-risk duties on model providers but does not require a dedicated corporate board AI committee. Meta’s planned committee adds a specific internal oversight mechanism beyond that provision, without replacing the provider’s legal accountability.
SB 53’s frontier-framework provisions address governance for implementing catastrophic-risk practices. They do not mandate a dedicated board AI committee, making Meta’s announced structure an additional planned governance measure.
Public preparedness reports
Meta promises preparedness reports for closed and open frontier releases, including assessments, evaluation methods, safeguards and deployment rationales. Reports must be updated promptly when circumstances materially change the prior risk assessment, with redactions explained where possible.
Article 53(1)(a) and (b) requires technical documentation and information for downstream providers, while Article 55 adds evaluation and systemic-risk duties. These provisions do not require a public, release-linked preparedness report explaining each deployment decision in the detail Meta promises.
SB 53’s transparency provisions require covered large developers to publish frontier frameworks and transparency reports addressing catastrophic-risk assessments and mitigation. Meta’s release-linked reports serve a comparable disclosure function, although publication under its own framework does not by itself establish statutory compliance.
Incident reporting commitments
Meta maintains a global incident-response program and commits to reporting critical incidents as appropriate. It also promises prompt preparedness-report updates when a major incident materially changes an assessment, but states no fixed authority-reporting deadline in this framework.
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, as appropriate, national competent authorities without undue delay. Meta’s discretionary wording does not express that mandatory reporting obligation.
SB 53’s critical-safety-incident provisions require reporting to California’s Office of Emergency Services within 15 days, with a 24-hour deadline for incidents posing an imminent risk of death or serious physical injury. Reporting as appropriate and updating a public preparedness report do not state those authority-notification deadlines.
The EU AI Act makes evaluation, systemic-risk mitigation, serious-incident reporting and cybersecurity legal duties for providers of general-purpose AI models with systemic risk under Article 55. Meta’s framework adds more specific controls for training-time containment, including sandbox validation and protected rollout logging, but limits its catastrophic-risk assessment to selected outcomes meeting four inclusion criteria. That narrower scope cannot substitute for the Act’s broader systemic-risk duties, and Meta’s commitment to report critical incidents as appropriate does not express the mandatory reporting standard found in EU law or California SB 53.
Source
https://research.meta.ai/static/meta-superintelligence-scaling-framework- Date on the page:
- "Version 2.1 | October 2026" in the running header of the 51-page PDF, and "Preface to Version 2.1 (October 2026)"; announcement post dated October 2, 2026.
- Source checked:
- opened and confirmed on 2026-10-30