Establishes a framework for managing and mitigating risks from frontier AI models at Meta, focusing on potential catastrophic outcomes. Involves threat modeling, risk assessment, and evaluations to determine access to and further development of the model
Analysis summaries, actor details, and coverage mappings were LLM-classified and may contain errors.
This is an internal corporate policy document establishing Meta's own framework for managing frontier AI risks. It contains voluntary commitments and internal governance processes rather than legally binding obligations with external enforcement.
The document has good coverage of approximately 10-12 subdomains, with strong focus on malicious actors (4.1, 4.2, 4.3), AI system security (2.2), competitive dynamics (6.4), governance failure (6.5), and AI safety failures (7.1, 7.2, 7.3). Coverage is concentrated in security, misuse prevention, and AI safety domains, with particular emphasis on catastrophic risks from frontier AI models.
This is an internal corporate policy document that governs Meta's own operations as an AI developer and information services company. The primary sectors governed are Information (where Meta operates) and Scientific Research and Development Services (Meta's FAIR Lab). The document does not regulate external sectors but rather establishes internal governance for Meta's frontier AI development activities.
The document comprehensively covers all stages of the AI lifecycle with particular emphasis on evaluation, deployment, and monitoring. It describes detailed processes for planning (threat modeling, reference class identification), evaluation and mitigation throughout development, deployment decisions based on risk thresholds, and post-deployment monitoring.
The document explicitly focuses on frontier AI models and systems, which are defined as highly capable general-purpose generative AI models. It does not use the terms 'foundation models' or 'GPAI' but the description aligns with these concepts. The framework addresses both open-weight releases and includes implicit compute considerations through capability-based thresholds rather than explicit FLOP thresholds.
Meta
Meta is the author and proposer of this framework, as evidenced by references to 'our Frontier AI Framework' and 'our wider AI governance program' throughout the document. Meta is developing frontier AI models and establishing internal governance processes.
Meta leadership teams; Multi-disciplinary review teams
Enforcement is conducted internally by Meta's leadership and multi-disciplinary teams who review risk assessments and make decisions about model development and release. There is no external enforcement body.
Meta internal teams; External subject matter experts; External experts
Monitoring is primarily conducted by Meta's internal teams throughout the development lifecycle, with involvement of external experts for specific evaluations and threat modeling exercises. The framework also references engagement with governments and the wider AI community.
Meta's internal research teams; Meta's product teams; FAIR Lab
The framework applies to Meta's own frontier AI development activities, including both research and product teams. It governs internal processes for developing, evaluating, and releasing frontier AI models.
11 subdomains (7 Good, 4 Minimal)