---
title: "Superintelligence Frameworks in 2026: A Sourced Guide"
slug: superintelligence-frameworks-2026
description: "What the main 2026 superintelligence frameworks actually say: Meta's scaling framework, Anthropic's RSP, NIST AI RMF, ISO/IEC 42001 and MAIM, with sources."
date: 2026-10-26
last_reviewed: 2026-11-02
status: published
type: knowledge-base
image: superintelligence-frameworks-2026
image_alt: "Glowing translucent spheres joined by fine threads across a dark blue network, an abstract picture of connected intelligence."
tags: [superintelligence, governance, safety frameworks, scaling policy, NIST AI RMF, MAIM]
---

# Superintelligence Frameworks in 2026: A Sourced Guide

**In short:** no single framework governs superintelligence. In 2026 the field is a layered set of lab scaling policies (Meta's Superintelligence Scaling Framework, Anthropic's Responsible Scaling Policy), management standards (NIST AI RMF, ISO/IEC 42001), a national-security proposal (Mutual Assured AI Malfunction) and a US executive order that changes federal terminology. They answer different questions, are written by different actors, and carry different weight. This guide separates what each one is, what it requires, and how well established the facts about it are.

Superintelligence here follows Nick Bostrom's definition: an intellect that greatly exceeds human cognitive performance in virtually all domains of interest. See [Artificial superintelligence](../vol06_asi/ASI-EXP-001_intelligence_explosion.md) for how the encyclopedia treats the transition from AGI, and the [glossary](/glossary.html) for terms such as recursive self-improvement and instrumental convergence.

## Quick reference

| Framework | Type | Owner | Current status (26 Oct 2026) | Evidence grade of the facts |
|---|---|---|---|---|
| Meta Superintelligence Scaling Framework | Lab scaling policy | Meta | v2.1, published 2 Oct 2026 | Established |
| Responsible Scaling Policy (RSP) | Lab scaling policy | Anthropic | v3.4, effective 8 Jul 2026 | Established |
| NIST AI RMF + Generative AI Profile (AI 600-1) | Risk-management framework | NIST (US) | Voluntary, in use | Established |
| ISO/IEC 42001 | Certifiable management standard | ISO/IEC | Published Dec 2023 | Established |
| Mutual Assured AI Malfunction (MAIM) | Deterrence proposal | Hendrycks, Schmidt, Wang | Proposal (Mar 2025), not adopted | Contested |
| Executive order "Inaugurating the Era of Super Intelligence" | Federal terminology order | US executive branch | Signed 29 Sep 2026 | Established |

The grades describe how well the *facts about each framework* are documented (what it says, when it was issued). They are not a verdict on whether the framework works. See the [evidence framework](/method/evidence.html) for the grading rules.

## What counts as a superintelligence framework

The term is used loosely. In practice it covers four jobs:

1. **Capability thresholds and scaling commitments.** When a model's measured capabilities cross a line, specific safeguards apply before further training or release. See [Dangerous capability and alignment evaluations](../vol08_safety_governance/SAFE-EVAL-001_dangerous_capability_and_alignment_evaluations.md).
2. **Risk management.** Repeatable processes for finding, measuring and treating risk, usable by any organization.
3. **Certification.** Auditable management systems that a third party can verify.
4. **Geopolitical strategy.** How states deter or manage a race to the most capable systems. See [International AI governance and compute governance](../vol08_safety_governance/GOV-INTL-001_international_ai_governance_and_compute_governance.md).

None of these is a technical definition of superintelligence, and none is a roadmap for building it. The underlying risk argument is covered in [The AI alignment problem](../vol08_safety_governance/SAFE-ALIGN-001_ai_alignment.md) and [Existential risk and catastrophic scenarios](../vol08_safety_governance/SAFE-XRISK-001_existential_risk.md).

## Frontier lab policies

### Meta Superintelligence Scaling Framework (v2.1)

On 2 October 2026 Meta [published an update](https://research.meta.ai/blog/developing-capable-models-responsibly) to its framework and renamed it the Meta Superintelligence Scaling Framework, version 2.1 ([full text](https://research.meta.ai/static/meta-superintelligence-scaling-framework)). It was previously the Advanced AI Scaling Framework (v2), which replaced the 2025 Frontier AI Framework. Meta says the update extends safety requirements to training and evaluation, clarifies considerations for open-weight releases, and adds a board-level AI committee to review future changes.

Meta ties the update to commitments made at the White House that week. The update also drew scrutiny: the watchdog group The Midas Project [documented changes](https://www.themidasproject.com/watchtower/meta-10022026) to cyber-capability thresholds that it describes as loosening earlier numeric standards. Treat that as one group's reading and check the redline against the text.

### Anthropic's Responsible Scaling Policy (v3.x)

Anthropic published the first [Responsible Scaling Policy](https://www.anthropic.com/responsible-scaling-policy) in September 2023, with AI Safety Levels (ASL) modelled loosely on biosafety levels. Version 3.0, a comprehensive rewrite, took effect on 24 February 2026 ([announcement](https://www.anthropic.com/news/responsible-scaling-policy-v3)); the version current at the time of writing is 3.4, effective 8 July 2026.

Two changes matter for anyone citing older descriptions. The policy no longer specifies escalating ASL tiers beyond those already implemented, and, according to an [analysis by GovAI](https://governance.ai/analysis/anthropics-rsp-v3-0-how-it-works-whats-changed-and-some-reflections), it dropped the earlier commitment to pause scaling if safeguards were not ready. Read the current text before relying on either description. This encyclopedia is built with Claude, a model from Anthropic, so check this section against the primary source.

### Other labs

Other developers publish similarly purposed documents under different names, such as safety or preparedness frameworks. METR keeps a [comparison of published frontier safety policies](https://metr.org/fsp) that is the easiest place to see versions and dates side by side.

## Standards and risk management

### NIST AI Risk Management Framework

The [NIST AI RMF](https://www.nist.gov/itl/ai-risk-management-framework) organizes risk management into four functions: Govern, Map, Measure and Manage. Its [Generative AI Profile (NIST AI 600-1)](https://doi.org/10.6028/NIST.AI.600-1) applies the framework to generative models, with twelve risk categories and more than two hundred suggested actions. Both are voluntary. They are the most widely referenced general-purpose backbone for enterprises and agencies, but they are not specific to superintelligence.

### ISO/IEC 42001

[ISO/IEC 42001](https://www.iso.org/standard/81230.html), published in December 2023, is the first international standard for an artificial intelligence management system. Its distinguishing feature is that an organization can be audited and certified against it, which the NIST framework does not offer.

## Government action

### The "Super Intelligence" executive order

On 29 September 2026 the US President signed an executive order titled "Inaugurating the Era of Super Intelligence" ([White House fact sheet](https://www.whitehouse.gov/fact-sheets/2026/09/fact-sheet-president-donald-j-trump-inaugurates-the-era-of-super-intelligence/)). It directs executive-branch agencies to use "Super Intelligence" and "SI" in place of "Artificial Intelligence" and "AI" in official, non-statutory documents.

It is a terminology order, not a safety framework. For its purposes, the new terms [mean the same technologies as the existing statutory definition of AI](https://iapp.org/news/a/ai-by-any-other-name-white-house-directs-federal-agencies-to-enter-the-era-of-super-intelligence), so it does not define or regulate superintelligence in the sense used in this guide. It also asks the President's science adviser to propose a federal definition within 60 days. The same day, technology executives signed a separate White House Accord on Super Intelligence ([CNBC report](https://www.cnbc.com/2026/09/29/trump-ai-super-intelligence.html)).

## Deterrence strategy

### Mutual Assured AI Malfunction (MAIM)

MAIM comes from the paper [Superintelligence Strategy](https://arxiv.org/abs/2503.05628) by Dan Hendrycks, Eric Schmidt and Alexandr Wang (March 2025). The argument: because sabotaging a rival's destabilizing AI project is relatively easy, through cyberattacks or potentially strikes on datacenters, states are already in a deterrence situation resembling nuclear mutual assured destruction. The paper pairs deterrence with nonproliferation and competitiveness.

MAIM is a proposal and a description of strategic incentives, not an adopted policy. Critics dispute the analogy, noting for example that AI development may be harder to observe and attribute than nuclear launches ([one critique](https://peterwildeford.substack.com/p/mutual-sabotage-of-ai-probably-wont)). Hence the "Contested" grade above.

## Not a framework: Safe Superintelligence Inc.

[Safe Superintelligence Inc. (SSI)](https://investor.nvidia.com/news/press-release-details/2026/Ilya-Sutskevers-Safe-Superintelligence-Inc--and-NVIDIA-Announce-Long-Term-Strategic-Partnership/) is a company, founded in 2024 by Ilya Sutskever and Daniel Levy, with a single stated goal: a safe superintelligence. It belongs in this discussion as an organizational approach (safety work integrated into core research rather than set out in a published policy), not as a framework others can adopt. On 27 July 2026 Nvidia and SSI announced a long-term partnership that includes an Nvidia investment and access to its Vera Rubin systems. See [AI safety institutes and research organizations](../vol10_people_organizations/ORG-SAFETY-001_ai_safety_institutes_and_research_organizations.md) for the wider landscape.

## Which one applies to you

| If you are... | Start with |
|---|---|
| Training or releasing frontier models | Your own lab's scaling policy, benchmarked against the METR comparison |
| An enterprise deploying AI | NIST AI RMF with the Generative AI Profile |
| An organization that needs third-party certification | ISO/IEC 42001 |
| A policy or national-security analyst | The MAIM paper and its critiques |
| A US federal contractor | The executive order's terminology requirements |

## Frequently asked questions

**What is the difference between AGI and superintelligence?**
AGI matches human-level performance across a wide range of intellectual tasks. Superintelligence greatly exceeds the best human performance in virtually all domains. See the [glossary](/glossary.html).

**Which framework should an enterprise use?**
The NIST AI RMF with its Generative AI Profile is the usual starting point. Choose ISO/IEC 42001 when formal certification is required.

**Does the "Super Intelligence" executive order regulate superintelligent AI?**
No. It changes terminology in federal documents and keeps the existing statutory definition of AI. Any new federal definition is still to be proposed.

**Is MAIM official US policy?**
No. It is a proposal from three authors, published as a paper in March 2025.

**Do these frameworks bind labs?**
The lab policies are voluntary and company-written, the NIST framework is voluntary, and ISO/IEC 42001 is voluntary but certifiable. None is a statute. For binding rules, see [Global AI governance and regulation](../vol08_safety_governance/GOV-LAW-001_ai_governance.md).

## Sources

- Meta: [Developing Capable Models Responsibly](https://research.meta.ai/blog/developing-capable-models-responsibly); [Meta Superintelligence Scaling Framework v2.1](https://research.meta.ai/static/meta-superintelligence-scaling-framework)
- Anthropic: [Responsible Scaling Policy](https://www.anthropic.com/responsible-scaling-policy); [RSP v3.0 announcement](https://www.anthropic.com/news/responsible-scaling-policy-v3)
- GovAI: [Anthropic's RSP v3.0: how it works and what's changed](https://governance.ai/analysis/anthropics-rsp-v3-0-how-it-works-whats-changed-and-some-reflections)
- METR: [Frontier AI safety policies](https://metr.org/fsp)
- The Midas Project: [Meta framework change log, 2 Oct 2026](https://www.themidasproject.com/watchtower/meta-10022026)
- NIST: [AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework); [AI 600-1](https://doi.org/10.6028/NIST.AI.600-1)
- ISO: [ISO/IEC 42001](https://www.iso.org/standard/81230.html)
- White House: [Fact sheet, 29 Sep 2026](https://www.whitehouse.gov/fact-sheets/2026/09/fact-sheet-president-donald-j-trump-inaugurates-the-era-of-super-intelligence/); IAPP: [analysis of the order](https://iapp.org/news/a/ai-by-any-other-name-white-house-directs-federal-agencies-to-enter-the-era-of-super-intelligence)
- Hendrycks, Schmidt, Wang: [Superintelligence Strategy](https://arxiv.org/abs/2503.05628)
- Nvidia: [SSI partnership announcement, 27 Jul 2026](https://investor.nvidia.com/news/press-release-details/2026/Ilya-Sutskevers-Safe-Superintelligence-Inc--and-NVIDIA-Announce-Long-Term-Strategic-Partnership/)

*Status: draft, reviewed 26 October 2026 against the primary sources above. Version numbers and policies in this area change often; check each source's current text.*