Trump Administration Eyes FINRA-Style Watchdog to Vet Frontier AI Models Before Public Release

The Trump administration is actively developing a plan to create an independent regulator for artificial intelligence modeled on the Financial Industry Regulatory Authority — the industry-funded watchdog that oversees broker-dealers and their registered representatives in the US securities industry. The proposal, first reported by Bloomberg on July 17, 2026, would establish a self-regulatory organization for frontier AI that tests models for safety risks and publishes standardized assessments, with oversight from the Securities and Exchange Commission.

The development represents a significant turn in the administration’s approach to AI governance. For the first eighteen months of Trump’s second term, the White House’s posture on AI regulation was almost uniformly focused on cutting red tape and accelerating American AI development — a deliberate contrast to what Republicans characterized as the Biden administration’s overly cautious, innovation-stifling approach. The FINRA proposal suggests that the administration has concluded that some form of structured safety evaluation is necessary to prevent the kind of catastrophic AI failures that could trigger far more aggressive congressional action.

What the Proposal Envisions

According to people familiar with the discussions, Treasury Secretary Scott Bessent has been the primary architect of the proposal within the administration. Bessent, who spent decades in the financial industry before joining the administration, is reportedly drawn to the FINRA model because it demonstrates how an industry with systemic risk potential can be regulated without creating a heavy-handed government bureaucracy.

The proposed structure would create a committee of independent technical experts — drawn from academia, national laboratories, and industry — to evaluate frontier AI models against a defined set of risk categories. The initial focus, according to people familiar with the discussions, would be on two areas where catastrophic misuse risk is highest: cybersecurity vulnerabilities that a model could enable or exploit, and biological threats, including models that could assist in the design or synthesis of pathogenic agents.

The organization would be funded by the AI industry — meaning companies like OpenAI, Anthropic, Google, Microsoft, xAI, and Meta would pay membership fees and assessments to support its operations, similar to how securities broker-dealers fund FINRA. This industry-funding structure is deliberately designed to avoid creating a new government agency that would require congressional appropriations and face the kind of budget and staffing instability that has plagued the US AI Safety Institute at NIST.

The SEC connection is less obvious at first glance. FINRA technically operates as a self-regulatory organization registered with and overseen by the SEC, but its day-to-day operations are independent of direct government control. The analogy for AI would be an organization that develops and enforces safety standards for frontier models, with the SEC providing a legal framework for enforcement if companies fail to comply — rather than having the SEC directly regulate AI model deployments.

The Problem It Is Trying to Solve

The proposal is a response to a genuine governance gap. Frontier AI models are increasingly deployed in contexts where failures could cause harm at scale — healthcare systems, financial markets, critical infrastructure, and potentially weapons development. But the United States currently has no systematic pre-deployment evaluation process for these systems.

The existing safety infrastructure is fragmented: the US AI Safety Institute at NIST conducts voluntary evaluations but has no enforcement authority. The Center for AI Standards and Innovation has signed evaluation agreements with Google DeepMind, Microsoft, and xAI, but these are voluntary partnerships with no standardized methodology. The Trump administration’s own AI Action Plan, signed in early 2026, called for “voluntary safety commitments” from AI developers — an approach that critics argued was inadequate given the pace of model capability development.

What a FINRA-style body would add is a standardized evaluation framework, an independent testing methodology, and — crucially — a compliance mechanism. Companies that develop frontier models above a certain capability threshold would be required to submit to evaluation before deployment, and the organization’s findings would be made public.

Industry and Expert Reaction

The proposal has drawn a notably warm reception from corners of the industry that might be expected to resist regulation. Google DeepMind CEO Demis Hassabis explicitly called for a new US-led global AI watchdog in a July 14 Axios interview, saying it was necessary “before year end” to address the risks emerging from frontier model development. Hassabis’s call was interpreted by many observers as a deliberate signal that at least some frontier AI labs would welcome a credible independent evaluation body.

The political logic for this apparent paradox is straightforward. Companies like Anthropic and Google DeepMind have invested heavily in internal safety research and evaluation infrastructure. A mandatory industry-wide evaluation process would impose compliance costs on all players, but it would also create a competitive dynamic where companies with strong safety cultures — and the evaluation results to prove it — are better positioned with enterprise customers, governments, and regulators. It is the classic dynamic in which incumbents with established compliance capabilities support regulation that raises barriers for newer entrants.

Not everyone is enthusiastic. Several AI startup founders and investors have pushed back on the proposal, arguing that mandatory pre-deployment evaluations would slow the pace of American innovation and create a bottleneck that Chinese labs — which face no equivalent restriction — could exploit. This concern is real, though proponents counter that the existing capability evaluation agreements suggest deployment timelines would not be significantly affected for well-prepared developers.

Parallel Tracks and Global Context

The FINRA proposal is not the only AI safety initiative moving through Washington. The Illinois AI Safety Measures Act, which requires annual independent third-party audits of frontier AI models, became the nation’s first state-level mandatory AI safety audit requirement earlier this month. Several other states are watching Illinois and considering similar legislation. Congress is reviewing the comprehensive Great American Artificial Intelligence Act, a bipartisan discussion draft introduced in June that would create a federal framework for AI governance.

Internationally, the FINRA proposal arrives at a moment when multiple governments are racing to establish credible AI governance frameworks. The EU AI Act’s first enforcement provisions took effect in February 2026, with the next tranche of requirements — including transparency obligations for general-purpose AI models — set for August 2. China has been actively building its own AI governance architecture, most recently through the 29-nation WAICO alliance announced earlier this week. And Demis Hassabis’s call for a US-led global AI watchdog suggests that at least some frontier lab leaders see American regulatory leadership as preferable to a fragmented global patchwork.

What Remains Unclear

The proposal is still in early development, and several critical details remain unresolved. What capability threshold triggers mandatory evaluation? How would the organization handle proprietary model weights and architectures during evaluation without creating competitive intelligence risks? What happens to a company that deploys a model before completing evaluation? And critically: how would a US-based evaluation regime interact with the regulatory requirements of the EU AI Act and emerging frameworks in other jurisdictions?

These are tractable problems — FINRA has resolved analogous questions for the securities industry over its 16-year history — but they will require substantial stakeholder engagement before a workable proposal can be sent to Congress or implemented through executive action.

What the emergence of this proposal makes clear is that the question in Washington is no longer whether frontier AI will be regulated, but how. The administration that campaigned on AI deregulation is now designing a regulatory body. The only remaining debate is whether the body it designs will be adequate to the task.