Frontier Model Forum
AI-generated text. This page was generated using artificial intelligence.
Frontier Model Forum (FMF) is an industry-supported nonprofit concerned with frontier AI safety and security. Its work includes identifying practices, supporting standards development, advancing safety research and sharing information among industry, governments, academia and civil society. The Forum states that it does not engage in lobbying. Its relevance to policy lies in the governance methods and technical practices it develops and discusses.[1]
Governance, membership and funding
Established in 2023, the Forum identifies itself as a section 501(c)(6) organization. An operating board comprising member representatives oversees it, and its published account names Chris Meserole as executive director. Membership fees fund its operations. This industry governance is relevant context for interpreting its reports as contributions from participating developers rather than statements by an independent regulator.[1]
The membership page reviewed in September 2026 lists Amazon, Anthropic, Google, Meta, Microsoft and OpenAI. Admission criteria include demonstrated frontier-model capability, published risk-management guidelines, safety-review processes and support for research and external evaluation. Members must also commit to participation and financial support for at least three years. These are the Forum's membership conditions, not a certification that every member's systems satisfy all applicable laws.[2]
Risk thresholds
The June 2025 report on risk taxonomy explains why frontier frameworks focus on selected severe risks, including biological and chemical threats, advanced cyber capabilities and certain autonomous behaviors. It distinguishes the point at which a capability requires additional scrutiny from the later determination that training or deployment is acceptable after safeguards are applied. Conflating these thresholds can obscure whether a developer has identified a hazard or actually justified proceeding.[3]
The report also distinguishes fixed historical risk baselines from assessments of the extra risk a new model adds to an evolving environment. A moving baseline may accommodate changes but also allow accumulated increases in risk. FMF presents this as an unresolved governance problem requiring further research and public discussion, rather than declaring a universally accepted level of risk.[3]
Capability assessments
The April 2025 assessment report compares three approaches. Relative assessments compare a model with previously examined systems; bottleneck assessments test abilities considered necessary for a harmful outcome; threat simulations examine a larger sequence of activity. The report stresses that these methods provide different kinds of evidence and have different demands on time, expertise and resources.[4]
Its organizational recommendations include separating assessment from development responsibilities, adequately resourcing evaluators and establishing accountability for decisions following testing. It also highlights uncertainty, appropriate baselines and the difference between knowledge tests and evidence of real-world capability. The report surveys emerging practices across the field; it is not a manual certifying one member's evaluation process.[4]
Independent and third-party assessment
The August 2025 report distinguishes confirmation of internal findings, testing with independently chosen methods and supplementation of a developer's expertise or capacity. Hiring an outside specialist to perform a developer-designed test serves a different function from an evaluator challenging the developer's assumptions with another method. The report treats access, expertise, rigor and actionable findings as central considerations.[5]
FMF expressly limits the report's scope: it does not cover all AI audits or regulatory-compliance assessments. It describes emerging practices rather than prescribed standards and recognizes variation among members. Its publications therefore should not be cited as proof that a mandatory, uniform industry assessment regime exists.[5]
Selected contributions
- Frontier Capability Assessments (April 2025).[4]
- Risk Taxonomy and Thresholds for Frontier AI Frameworks (June 2025).[3]
- Third-Party Assessments (August 2025).[5]
Related articles
- Policy
- United States policy on catastrophic AI risk
- OpenAI
- Anthropic
- Google DeepMind
- Partnership on AI
References
- ↑ 1.0 1.1 1.2 1.3 1.4 Frontier Model Forum, About. Accessed September 12, 2026.
- ↑ Frontier Model Forum, Membership. Accessed September 12, 2026.
- ↑ 3.0 3.1 3.2 Frontier Model Forum, Risk Taxonomy and Thresholds for Frontier AI Frameworks, June 18, 2025. Accessed September 12, 2026.
- ↑ 4.0 4.1 4.2 Frontier Model Forum, Frontier Capability Assessments, April 22, 2025. Accessed September 12, 2026.
- ↑ 5.0 5.1 5.2 Frontier Model Forum, Third-Party Assessments, August 4, 2025. Accessed September 12, 2026.