The White House is negotiating with Anthropic to establish AI security rules, with National Economic Council director Kevin Hassett comparing the approach to FDA drug safety evaluations — and signaling a potential executive order that would end the Trump administration’s hands-off AI stance.
🔍 THE BOTTOM LINE
The US government is quietly building a pre-deployment safety regime for frontier AI models, centered on CAISI (the Center for AI Standards and Innovation inside NIST). Forty evaluations are done. Five major labs — Anthropic, OpenAI, Google DeepMind, Microsoft, and xAI — have signed on. But CAISI has 30 staff and $30 million since 2024. Singapore spends more. If this becomes the global standard, Five Eyes partners including New Zealand will be expected to align with a regime that is, right now, structurally underweight.
What Changed
The shift was triggered by Anthropic’s Mythos model demonstration, which showed frontier AI could find and exploit decades-old vulnerabilities in widely used software. According to Federal News Network’s reporting, Hassett told Fox Business: “We’re studying possibly an executive order to give a clear road map to everybody about how this is going to go and how future AI that also potentially create vulnerabilities should go through a process so that they’re released in the wild after they’ve been proven safe, just like an FDA drug.”
That is a pivot. The Trump administration spent its first months dismantling Biden-era AI oversight, framing regulation as a drag on innovation. Now the same administration is comparing AI models to pharmaceuticals — the most heavily regulated product class in the US economy.
CAISI has already signed pre-deployment evaluation agreements with Google DeepMind, Microsoft, and xAI, adding to existing arrangements with Anthropic and OpenAI. Forty evaluations have been conducted, including on unreleased models. Commerce Secretary Howard Lutnick designated CAISI as industry’s “primary point of contact” for AI testing.
The Funding Gap
Here is the contradiction. The America First Policy Institute — a Trump-aligned think tank — called CAISI “chronically underfunded” in a recent issue brief. Approximately 30 total staff. $30 million since 2024. Less than comparable AI centers in Canada and Singapore. The brief recommended $50-100 million in annual funding.
The Federation of American Scientists went further, proposing an annual operating budget of up to $155 million plus $155-275 million in setup costs for high-security compute facilities. CAISI is being asked to evaluate the most powerful technology in human history with the staffing level of a mid-sized law firm.
This matters because the US is setting the de facto global standard. If CAISI’s evaluations become the gatekeeper for frontier model deployment — through an executive order or otherwise — then every country that wants access to those models will need to accept the CAISI framework as the baseline. That includes New Zealand.
NZ Angle
New Zealand is a Five Eyes intelligence partner. When the US sets security standards for dual-use technology, alignment is not optional — it is the price of interoperability. The Mythos rollout to Australia already demonstrated that Anthropic’s security-first models reach Five Eyes partners through a gated process, not open market access.
If the White House issues an executive order mandating CAISI pre-deployment evaluations, NZ’s National Cyber Security Coordinator and the Government Communications Security Bureau (GCSB) will need to assess what alignment looks like. Does NZ accept CAISI evaluations as sufficient? Does it run its own parallel process? Or does it simply defer to the US framework and lose any independent assessment capability?
The answer matters because NZ has no equivalent of CAISI. The GCSB’s cyber capability is focused on threat intelligence and incident response, not frontier model evaluation. If the US standard becomes the global standard, NZ’s options are alignment or irrelevance.
This is also a story about export controls and sovereign AI. The White House already treated Anthropic’s models as munitions under the EAR. Now it is building a safety regime around them. The two tracks — export control and domestic safety regulation — are converging on the same outcome: frontier AI is being treated as a regulated industry, and the US government is the regulator.
The Other Side
Not everyone agrees the FDA model works for AI. The comparison implies a static product that can be tested once and approved. Frontier models are updated continuously, fine-tuned by user interactions, and deployed in contexts their creators cannot fully predict. An FDA-style approval process designed for a drug that stays chemically identical for years may not fit a model whose behavior shifts week to week.
There is also the guardrail-paradox problem. As the OALABS breach analysis showed this week — where a low-skilled attacker in Ethiopia used Claude and Codex to breach 14 companies — the framing that bypasses AI safety guardrails (“authorized red team exercise”) is the same framing used by legitimate security researchers. Broader refusals hurt defenders more than attackers, who can simply use less restrictive models. An FDA-style process that makes frontier models harder to access for legitimate researchers without addressing the framing problem would make cybersecurity worse, not better.
The Bigger Picture
The Trump administration’s AI policy has lurched from “remove all Biden-era rules” to “build an FDA for AI” in the space of three weeks. The trigger was not a theoretical risk paper — it was a live demonstration that Mythos could find real vulnerabilities in real software. When the capability is proven, the politics shifts.
The question for NZ is whether we are watching. The EFF’s constitutional challenge to the Anthropic export controls shows the US domestic debate is heated. But the rest of the Five Eyes is not in that debate. We are the recipients of the outcome, not participants in the process.
❓ FAQ
Will this executive order actually happen? Hassett said the White House is “studying” it. That is the precursor language. The Trump administration has shown it prefers executive action over Congressional legislation. Given the Mythos demonstration and the CAISI agreements already in place, an executive order is the path of least resistance.
What does CAISI actually test? Pre-deployment evaluations of frontier models — checking for capabilities that could enable cyberattacks, biological weapon design, or other dual-use applications. Forty evaluations have been conducted so far, including on unreleased models. The specifics are classified.
How does this affect NZ companies using AI? Short term, nothing changes. Medium term, if CAISI evaluations become a precondition for deployment, NZ companies using US-built frontier models (Claude, GPT, Gemini) will be using models that passed a US government safety gate. The question is whether NZ accepts that gate as sufficient or builds its own.
Is the FDA comparison accurate? Partially. Drugs are static molecules. AI models are continuously updated systems that behave differently in different contexts. The FDA model works for approving a specific version of a drug. It is less clear how it works for a model that changes behavior based on how millions of users interact with it.
🔍 THE BOTTOM LINE
The US is building an AI safety regime with 30 staff and less funding than Singapore. The ambition is FDA-scale. The capacity is not. For New Zealand, the signal is clear: the hands-off era is ending, and the standard being set is one we will be expected to follow — not one we helped design.
