Governance document

CA-D-000903

Meta Frontier AI Framework

Key Facts

What does Meta have in place to keep risks within acceptable levels?

Meta has processes in place to keep risks within acceptable levels, including applying mitigations.

If catastrophic outcomes are enabled by technological advances, what does Meta's framework do?

Meta's framework focuses on evaluating whether catastrophic outcomes are enabled by technological advances and, if so, identifying ways to mitigate those risks.

What does Meta define risk thresholds based on?

Meta defines risk thresholds based on the extent to which its models facilitate threat scenarios.

What potential outcomes does Meta's framework identify?

Meta's framework identifies potential catastrophic outcomes related to cyber, chemical, and biological risks that Meta strives to prevent.

What does Meta strive to prevent?

Meta's framework identifies potential catastrophic outcomes related to cyber, chemical, and biological risks that Meta strives to prevent.

What potential outcomes does Meta's framework identify?

Meta's framework identifies potential catastrophic outcomes related to cyber, chemical, and biological risks that Meta strives to prevent.

What does Meta strive to prevent?

Meta's framework identifies potential catastrophic outcomes related to cyber, chemical, and biological risks that Meta strives to prevent.

What does Meta commit to continuing to do?

Meta commits to continuing to evolve and refine its approach as AI technology progresses.

What does Meta's open source approach enable it to do?

Meta asserts that its open source approach enables it to learn from the broader community's independent assessments of its models' capabilities, which helps it better anticipate and mitigate risk.

What does Meta's framework take into account in addition to evaluating catastrophic risk?

Meta's framework takes into account the benefits of the technology it is developing in addition to evaluating catastrophic risk.

Stay ahead of the changes

Track Meta and get the diff the day its terms change.

Monitor Meta + up to 20 platforms — $4.99/mo →

Summary

This document describes how Meta approaches the responsible development of advanced AI, including how it identifies and tries to prevent serious harms in areas like cybersecurity, chemical risks, and biological risks. Meta uses threat modeling and external expert input to assess risks, and factors in the benefits of its technology alongside those risks. The framework is not final — Meta commits to updating it as AI technology develops.

Analysis

This document establishes Meta's internal framework for identifying, evaluating, and mitigating catastrophic risks arising from frontier AI development, with preventive focus scoped to cyber, chemical, and biological risk domains. Meta defines risk thresholds based on the degree to which its models facilitate threat scenarios, and maintains operational processes — including mitigations — to keep risks within acceptable levels. The framework incorporates both catastrophic risk assessment and technology benefit considerations, conducts structured threat modeling exercises with external experts where necessary, and relies in part on community assessments of model capabilities enabled by Meta's open-source release approach. Meta commits to evolving and refining the framework as AI technology progresses.

What this means for you

For an individual, this document describes Meta's internal risk-management commitments rather than establishing direct user rights or obligations. Meta's framework targets catastrophic outcomes in cyber, chemical, and biological domains, with risk limits set based on how much its models could facilitate harmful scenarios. Meta has stated processes to apply mitigations when risks exceed acceptable levels. The document does not establish specific user-facing opt-out mechanisms, preference settings, or contact windows.

Institutional Analysis

Stay ahead of the changes

Institutional analysis available with Insight

Which mapped governance frameworks each document engages, tied to the specific provisions that engage them.

Monitor Meta + up to 20 platforms — $4.99/mo → Get the research letter, free

15 important changes detected

16 versions captured · Last updated: August 2026

August 19, 2026

low

What changed Meta's Frontier AI Framework detected changes on August 19, 2026 that involved adding three sentences and modifying one sentence related to workforce development and partnerships. The specific content changes involved adding references to 'No Prior Experience Necessary' and 'A Career on the Other Side' alongside existing partnership announcements with NABTU. The operational significance of these changes appears limited to content organization and messaging related to Meta's workforce initiatives rather than core AI governance or model risk framework modifications.

Why this matters This change does not materially affect the terms or policies consumers operate under. The updated content adds messaging about Meta's workforce development initiatives and partnership announcements but does not modify governance frameworks, model risk procedures, or consumer-facing rights and obligations.

View full change record →

August 13, 2026

low

What changed Meta's Frontier AI Framework was revised in an update detected on August 13, 2026. A single sentence was modified to replace a reference to 'Introducing WhatsApp Web Calling, Plus More New Features' with 'NABTU and Meta Announce New Partnership to Invest In Skilled Trades for the AI Era'. This change updates the related news section but does not materially alter the framework's governance, risk assessment, or operational policies.

Why this matters This change updates a reference in the Frontier AI Framework's related news section but does not alter the framework's substantive governance provisions, risk assessment methodologies, or operational policies. No action is required on the part of users or organizations.

View full change record →

August 11, 2026 low

Meta's Meta Frontier AI Framework detected a change in an update detected on August 11, 2026. One sentence in the document's 'Recent News' section was modified: the article headline 'We're …

View change record →

August 7, 2026 low

In an update detected on August 7, 2026, Meta's Frontier AI Framework modified its 'Recent News' section. The change removed the headline 'Announcing the AI Glasses Impact Grant Recipients: Helping …

View change record →

August 5, 2026 low

Meta updated the Meta Frontier AI Framework on August 5, 2026 by adding a new recent news headline, 'We're Upgrading Your WhatsApp Group Chats', to the document's news section. This …

View change record →

July 29, 2026 low

Meta updated the related news section of their Meta Frontier AI Framework in an update detected on July 29, 2026. The document removed references to 'Meta AI Doesn't Just Think, …

View change record →

July 28, 2026 low

Meta's Frontier AI Framework was updated to replace a reference in the related news section. The previous reference to 'Connecting Real People on Facebook' was removed, and replaced with 'Announcing …

View change record →

July 25, 2026 low

Meta updated the 'Recent News' section of their Meta Frontier AI Framework document on July 25, 2026, replacing references to WhatsApp features, Threads parental tools, and Meta AI alerting with …

View change record →

July 23, 2026 low

Meta updated the Meta Frontier AI Framework on July 23, 2026, revising the related news section to replace a reference to 'Teachers and Local Businesses Win as Meta Expands Louisiana …

View change record →

July 22, 2026 low

Meta's Meta Frontier AI Framework was updated on July 22, 2026. The change modified the 'Recent News' section of the document, replacing a reference to 'Breaking Ground on Meta's First …

View change record →

July 18, 2026 low

Meta updated its Frontier AI Framework on July 18, 2026 by modifying a single sentence in the document's Recent News section. The prior version listed recent articles about Meta's AI …

View change record →

July 17, 2026 low

Meta's Frontier AI Framework removed one news item from its Related News section, replacing a generic reference to 'Introducing Muse Image: Image Generation Built for Your World' with a specific …

View change record →

July 14, 2026 low

In an update detected on July 14, 2026, Meta's Frontier AI Framework changed a single sentence in the 'Recent News' section. The prior version listed 'It's Time to Reserve Your …

View change record →

July 9, 2026 low

On July 9, 2026, Meta updated the navigation and recent news section of its Frontier AI Framework document. The change removed a reference to 'Inside One of Meta's Data Centers' …

View change record →

July 8, 2026 low

Meta updated the recent news section of its Frontier AI Framework document on July 8, 2026, replacing references to partnerships and product launches with new AI-focused announcements. The previous version …

View change record →

Featured, High severity

Apply mitigations to maintain acceptable risk Compare →

Meta asserts the existence of operational processes designed to maintain risks within acceptable levels, with mitigation as one named mechanism.

Added July 10, 2026 Standard Seen across 277 platforms
Assess technological enablement of harms Compare →

Meta's risk assessment is conditioned on a causal link between technological advancement and catastrophic harm — mitigation steps are triggered only where that link is identified.

Added July 10, 2026 Unique
Define risk thresholds for model facilitation Compare →

The metric Meta uses to set risk thresholds is the degree to which its models facilitate threat scenarios, meaning model behavior directly determines where risk limits are set.

Added July 10, 2026 Unique
Evaluate catastrophic risk before model release Compare →

Meta has defined specific risk domains — cyber, chemical, and biological — as the scope of catastrophic outcomes its framework targets, establishing the boundaries of its preventive focus.

Added July 10, 2026 Standard Seen across 263 platforms
Identify catastrophic cyber bio outcomes Compare →

Meta has explicitly scoped its catastrophic risk prevention efforts to three domains — cyber, chemical, and biological — which defines the boundaries of its framework's protective reach.

Added July 10, 2026 Standard Seen across 282 platforms
Open source release as default approach Compare →

Meta frames open sourcing as a non-optional imperative tied to national interest, which provides the stated justification for its open source release approach.

Added July 10, 2026 Standard Seen across 257 platforms

Featured, Medium severity

Commit to evolve framework over time Compare →

The commitment signals that Meta's current framework is not static, meaning the standards and practices it describes may change over time.

Added July 10, 2026 Standard Seen across 289 platforms
Community independent assessment of models Compare →

Meta's claimed risk-mitigation process partially depends on external community assessments rather than solely internal evaluation.

Added July 10, 2026 Standard Seen across 273 platforms
Consider benefits alongside catastrophic risk Compare →

Risk evaluation under Meta's framework is not one-sided; potential benefits are explicitly factored in alongside catastrophic risk, which shapes how risk thresholds and mitigations are determined.

Added July 10, 2026 Unique
Threat modeling with external experts Compare →

Meta's catastrophic risk assessment is an active, structured process that involves external expertise, not solely internal review.

Added July 6, 2026 Unique

Featured, Low severity

Share framework to advance AI evaluation research Compare →

Meta frames the publication of this framework as intended to contribute to broader industry and research conversations, not solely as internal policy documentation.

Added July 10, 2026 Standard Seen across 273 platforms