Executive Order 14409 (Promoting Advanced AI Innovation and Security)

Executive Order 14409 of 2 June 2026, "Promoting Advanced Artificial Intelligence Innovation and Security," signed by President Donald Trump (Federal Register, published 2026-06-05). It is the standing US instrument behind the June-2026 frontier-model release controls — the policy frame the Anthropic ban lacked a name for and that the GPT-5.6 gating revealed.

What it requires — corrected 2026-08-30, and the correction is the whole page

This page said the order compels something. It does not. The Congressional Research Service's own explainer (crs-eo-14409-explained, IF13268 v2, 9 July 2026, T1) states that companies "are asked to participate in a voluntary review process," and that the order's provisions "stop short of formal licensing or requiring preclearance." The 30-day window is real; the obligation is not.

Superseded text, kept per record-don't-overwrite: "30-day pre-release submission. AI companies must submit new model releases to the government 30 days before release… A standing review gate, not a one-off. Where an export designation says 'you may not ship,' EO 14409 builds a recurring pre-clearance pipeline for every frontier release." Sourced from trade-press reporting in June; not checked against the order or a primary analysis until now.

What the order actually does:

A voluntary 30-day window, offered rather than imposed, and in CRS's phrasing it precedes release "to other trusted partners" — critical-infrastructure operators being the example — though the same report elsewhere describes it as preceding public release. Both framings are in the one document and both are recorded.
No definition of "covered frontier model." The order does not supply one. Defining it was itself a deliverable, due 1 August 2026 out of a classified benchmarking process.
Nine tasks, all of them on agencies, not on companies — CNSS and DOD system hardening, CISA Binding Operational Directives, a Treasury-led AI cybersecurity clearinghouse, grant identification, the benchmarking process, a voluntary model-disclosure framework, OPM hiring pathways, and DOJ prosecution priorities. Full table in crs-eo-14409-explained.
Funded from existing appropriations, with Treasury not having requested FY2027 money for the clearinghouse it is supposed to build.

The stated rationale quoted before still stands: advanced AI "make[s] the nation stronger" but "introduce[s] new national security considerations that require coordinated action."

Why the error was easy to make, and worth keeping visible. A voluntary framework and a mandatory gate look identical from the outside when every lab participates and two of them are publicly restricted in the same fortnight. The distinction only becomes visible in the primary text — and it is the distinction that decides whether the section below is describing enforcement or coincidence.

The "enforcement record" — and CRS does not call it that

Read this section against the correction above. A voluntary framework has nothing to enforce, and crs-eo-14409-explained lists the Anthropic action in contrast to the EO rather than as an instance of it, inviting Congress to weigh the EO's voluntary assessment "especially in contrast to stronger government direction or regulation. It may also choose to consider the Administration's other approaches, such as requiring Anthropic to suspend access to a frontier AI model showing advanced cyber capabilities."

So there are two readings of the same fortnight and this wiki now holds both. Ours, written in June: a standing gate producing two beats. CRS's, in July: a voluntary framework, plus separate and stronger ad-hoc actions taken under other authority. The second is better sourced and is the one to prefer; the first is kept because the sequencing that produced it is real and unexplained — two frontier labs restricted within a fortnight of an order about frontier-model security is a coincidence that still wants an account.

The two beats, as recorded in June:

Anthropic — Fable 5 / Mythos 5 (≈12 June 2026) — suspended worldwide over a contested cyber-misuse/jailbreak claim (david-sacks the named framer).
OpenAI — GPT-5.6 (reported 25–26 June 2026) — release limited to government-approved partners, approved customer-by-customer during preview. Corroborated from the provider side: the GPT-5.6 system card rates the model High in cyber and bio/chem (preparedness-framework) and states "government coordination was requested before broader deployment" — the lab confirming the gate this EO formalizes.

So the same regime produced an outright suspension in one case and graduated access-gating in the other — the gate has a dial, not just an on/off switch.

The first completed cycle — and the missing half

The Anthropic models came back online on 1 July 2026 after a 20-day review, the first time the regime has run a full ban → review → re-release loop. What that cycle reveals is that EO 14409 is only half a process: it mandates pre-release submission but names no approval procedure. The re-clearance ran as an ad hoc, multi-agency crash review — the Center for AI Standards and Innovation and the NSA rejected Anthropic's first fixes, agency heads signed off gradually, Commerce coordinated — with no published criteria, threshold, or timeline. The Axios account's own bottom line is that a "framework for approving future models with a clear inclusive process, transparency standards and timelines" still doesn't exist. So the standing instrument has a defined entry (submit 30 days out) and an undefined exit (approval by discretion) — the governance gap the next iteration has to close.

Where it sits on the wiki's instrument map

It is the formalization of export control turned inward at domestic labs — but via administrative pre-review rather than a chip-style trade ban, a distinct mechanism.
It sharpens the contradiction inside the US cell: us-ai-policy frames the administration as deregulatory (the Dec-2025 preemption EO clearing state AI laws away), yet EO 14409 is a hard, discretionary security gate on release. The same administration deregulates conduct and pre-clears frontier releases — risk-based in spirit, access-denial in mechanism.
Efficacy is contested (cyber-export-control-history): a release gate slows spread without undoing capability, and practitioners read it as accelerating the move to self-hostable open-weight / Chinese models — the security goal and the market reaction running against each other.

Caveat

The EO text and date are verifiable (Federal Register 2026-06-05). What's reported (secondhand) is how it's being applied to specific models — the worldwide Anthropic suspension and the per-customer GPT-5.6 mechanism come via trade press and staff Q&As, not official designation documents. Treat the instrument as solid, the case-by-case enforcement detail as a contested snapshot.

anthropic-export-ban-2026 · openai-gpt56-access-restriction · gpt56-system-card · preparedness-framework · export-controls-on-ai · · anthropic-mother-of-all-risk-factors us-ai-policy · cyber-export-control-history · risk-based-regulation · ai-governance · synthesis