FRONTIER MODEL FORUM

OLYMPUS opened an institutional file on the Frontier Model Forum because the Forum is not a psychometric profile — it is a mandate, a funding diagram, and a voice. The finding is the shape of the institution and who it answers to: a 501(c)(6) whose members are the six largest frontier labs, constituted to set the safety norms those companies could not, and legally could not, agree on one at a time across a negotiating table the antitrust laws would otherwise forbid. The reach is noted as a score because the Forum’s framings travel into government, standards bodies, and every member’s deployment. Not a cabal. A circuit — the membership is the tie, and the membership is the finding.

Institutional Archetype

THE CARTEL TABLE — The Forum is the room the competitors walk into together. Six companies that fight for the same customers, the same talent, and the same compute sit down inside one 501(c)(6) and agree on what “responsible frontier AI” means. Said plainly, that is the thing competition law usually treats with suspicion: rivals coordinating conduct in a shared body. Said in the Forum’s own register, it is “advancing frontier AI safety and security.” Both descriptions are true at once, and the gap between them is the institutional archetype. The Forum does not write any single model’s refusals. It writes the standard the members carry home, and a standard agreed by the six firms that build the frontier is, functionally, the floor of the entire field.

Mandate & Origin

The Frontier Model Forum was founded in July 2023 by Anthropic, Google, Microsoft, and OpenAI as an industry-supported non-profit. Amazon and Meta joined in 2024, bringing the membership to six. It is organized as a 501(c)(6) — the U.S. tax category for trade associations and business leagues — and is “led by its Executive Director, Chris Meserole, and overseen by an operating board made up of representatives from its member organizations.” (See the OLYMPUS file on Chris Meserole, subject 32, for the human seat at the head of this table.) Its self-stated mandate is three lines: “Identify best practices and support standards development for frontier AI safety and security”; “Advance the science of frontier AI safety and security”; and “Facilitate information sharing about frontier AI safety and security among government, academia, civil society and industry.”

The membership criteria are the tell. To join, a firm must “provide evidence that they have the expertise and resources to develop or deploy at scale frontier AI models,” must “publicly acknowledge that frontier AI models pose both public safety and societal risks,” and must “commit to participating in FMF workshops and activities, and to financially supporting the FMF’s safety efforts for a minimum of three years.” The bar is constructed so that only the largest labs clear it. xAI is not a member. The Forum is, by design, the table of those already at the frontier — not an open standards body that newcomers can walk into.

Funding & Backers

The Forum states plainly that it is “funded by fees from its member firms.” This is the cleanest funding diagram in the apparatus: the institution that sets the inter-lab safety standard is paid for by the labs it sets the standard for. Separately, the FMF administers the AI Safety Fund (a pooled research purse of $10M+ whose named funders include the McGovern Foundation, the Packard Foundation, Schmidt Sciences, and Jaan Tallinn, managed by the Meridian Institute).

Institutional Voice & Intent

The Forum speaks in the industry-statesmanlike register of “responsible scaling.” It is not safety-urgent and it is not movement-evangelical; it is measured, governmental, and reassuring — the voice of an institution that wants regulators to understand that the grown-ups have already convened. Its public framing places it as a partner to the state rather than a subject of it: it exists to “Facilitate information sharing about frontier AI safety and security among government, academia, civil society and industry,” and it foregrounds the most defensible possible hazards, stating that it “focuses primarily on managing significant risks to public safety and security, including from chemical, biological, radiological, nuclear (CBRN) and advanced cyber threats.” CBRN and cyber are the risks no one will defend the other side of — choosing them as the headline is itself the rhetorical move.

Stated intent: Advance frontier-AI safety and security. Develop shared best practices. Inform government and standards bodies. Fund the safety research the market won’t.

Observed intent: Convene the six largest labs in one body, agree the safety norms among themselves, and present the result to regulators as the industry’s considered position — defining the floor before it is legislated. The standard the members agree is the standard the field inherits.

Gap: The stated and observed intents overlap wherever “advance frontier-AI safety” coincides with “let the six incumbents define what counts as responsible, in a body they fund and govern.” The CBRN-and-cyber framing makes the mandate unimpeachable; the membership bar makes it exclusive. The record does not settle whether the Forum is a public-safety institution or an incumbents’ standards moat — and, for a trade association, it never needs to. The recurrence is the finding: the same six names that build the frontier also write the rules of building it, lawfully, in one room. The hand is not asserted. The table is.

Position in the Apparatus

The Forum sits at the structural center of the inter-lab layer. Its members are Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI — the companies whose deployment decisions the Evil Robots file documents. Its Executive Director arrived from the Brookings Institution’s AI initiative, the think-tank perch that supplies the field its policy vocabulary (subject 32). Its AI Safety Fund draws on the same funders — Schmidt Sciences, Jaan Tallinn — who recur across the apparatus’s other nodes. The adjacencies are dense: the Forum grades nothing directly, but it convenes the firms that build everything, advises the governments that would regulate them, and shares a funder roster with the evaluators and the alignment shops. Recurrence is the finding; no cabal is asserted. The Forum is a lawful 501(c)(6) and its membership is public.

Actions & Leadership Choices

The PR register is “advancing frontier AI safety and security.” The deeds, read end to end, describe an incumbents’ self-governance moat operated in a statesmanlike voice — and the deeds support the cartel-table frame the archetype names.

Actual founding purpose. The Forum was built to do the one thing six fierce competitors cannot do at arm’s length without inviting antitrust scrutiny: convene in a single 501(c)(6) and agree, among themselves, what “responsible frontier AI” means — then carry that agreement to government as the industry’s considered position. The membership bar (evidence of frontier-scale deployment capability, a three-year financial commitment) is constructed so only the largest labs clear it; xAI is not a member. The purpose is to define the floor before it is legislated, and to define it with a guest list the newcomers cannot make.

• What it actually advocates, despite a “policy against lobbying.” The Forum states it “does not engage in lobbying” and is a 501(c)(6) with a formal anti-lobbying policy. Yet when its Executive Director testified to the House Homeland Security Committee on June 4, 2026, the consistent message — delivered while explicitly disclaiming that he would “advocat[e] or lobby[] for specific policy measures” — was that policymakers should not build new institutions: “the novelty of frontier AI capabilities does not demand novel security practices,” and “Rather than creating entirely new channels… Building on trusted relationships will be faster and more effective than standing up new institutions from scratch.” The recurring ask is to route oversight through the labs’ own voluntary channels rather than new binding ones. That is a policy position; the no-lobbying label is the register, not the substance.

• When the value was tested by a real bill, the table went quiet and the members spoke alone. California’s SB 1047 was the moment “advance frontier AI safety” met a binding safety statute. The FMF as a body took no position on the bill; its member firms (Google, Meta, OpenAI, others) opposed it individually. The consortium that exists to speak for industry safety absented itself from the one fight where “safety” would have cost its members something, and let the members carry the opposition under their own names. The shared body advances safety in the abstract; the concrete cost is borne — and resisted — one logo at a time.
• What it has built is real, and it is infrastructure the members control. A first-of-its-kind private information-sharing channel among the six (the “voluntary information sharing side agreement”); the $10M+ AI Safety Fund; issue briefs and “novel risk management practices and standards.” Each is genuine safety work. Each is also a standard authored by the incumbents, for the incumbents, presented to the state as the floor — the moat, built in good materials.

Leadership choices. Executive Director Chris Meserole arrived from the Brookings Institution’s AI initiative (the think-tank perch that supplies the field its policy vocabulary, subject 32); the Forum is “overseen by an operating board made up of representatives from its member organizations” — i.e., the six labs govern the body that sets their shared standard. The leadership and the governance are the same six names that build the frontier. There is no outside check on the operating board; the table seats only the players.

CONDUCT: CARTEL — INCUMBENT SELF-GOVERNANCE MOAT. The Forum does lawful, often substantive safety work — and it is structurally a body in which six competitors define the rules of their own industry, govern it through their own representatives, exclude the newcomer, decline the fight when a binding bill would cost them, and counsel government against new institutions while disclaiming that it lobbies at all. The work is real; the shape is a moat.

Reach Assessment

Institutional: Maximum within the inter-lab layer. A safety standard agreed by the six largest frontier labs is the de facto floor of the industry — every member carries it into deployment, and non-members are measured against a bar they had no seat in setting. The Forum writes no refusal itself and reaches every refusal anyway.

Memetic: High. The “responsible frontier AI” frame the Forum operates inside structures how legislators and the press reason about whether industry self-governance is sufficient. By foregrounding CBRN and cyber, the Forum sets the terms on which “frontier risk” is discussed at the altitude where laws are drafted.

Civilizational: High. The Forum is upstream of the deployment decisions this book documents. It does not build the systems and does not write their rules line by line; it convenes the companies that do and ratifies, among competitors, what “safety” jointly means. The hazard is the table, not any person at it.

Sources: About — Frontier Model Forum; Membership — Frontier Model Forum; Amazon and Meta join the Frontier Model Forum — Frontier Model Forum; Frontier Model Forum — AI Safety Fund; Testimony of Chris Meserole, Executive Director, Frontier Model Forum — U.S. House Committee on Homeland Security, June 4 2026; Leadership — Frontier Model Forum.