CRS — "Controlling Advanced Artificial Intelligence: Executive Order 14409 Explained"
Congressional Research Service In Focus IF13268, version 2, dated 9 July 2026. Chris Jaikaran
(cybersecurity policy), Laurie Harris (science and technology policy) and Kelley M. Sayler (advanced
technology and global security). Three pages, read from the PDF with pypdf after congress.gov
returned 403 to a plain fetch.
T1, and the highest-quality source this spoke holds on the US instrument. CRS is nonpartisan staff writing for Members of Congress about a document Congress did not pass — no vendor, no trade press, no advocacy. It supersedes this wiki's earlier reading of eo-14409, which was built from reporting.
The correction, and it is the load-bearing sentence
eo-14409 has said since June that "AI companies must submit new model releases to the government 30 days before release," and called the order "a recurring pre-clearance pipeline." CRS says the opposite, twice:
"Companies developing advanced AI models are asked to participate in a voluntary review process that gives the government a window of 30 days to examine new systems before such models are released to 'other trusted partners' (e.g., critical infrastructure companies)."
"The E.O.'s cybersecurity provisions aim to expand voluntary national security oversight of advanced AI models while stopping short of formal licensing or requiring preclearance."
Three things were wrong at once, and they compound:
- Not mandatory. Participation is requested. There is no obligation and no penalty named.
- Not a gate on public release in the passage above — the 30 days precede release to other trusted partners, critical-infrastructure operators being the example given. (The report's later discussion does say the window aims to give agencies time to probe models "before public release," so the two framings sit side by side in the same document and are recorded that way rather than resolved.)
- The order does not define "covered frontier model." CRS states this plainly under Considerations for Congress. The definition was itself a deliverable, due within 60 days (1 August 2026) via a classified benchmarking process.
CRS also names the gap the voluntary structure creates: coverage fails "if major developers decline to participate, criteria for covered models are narrowly tailored, or the length of review is insufficient."
What the order actually directs, with deadlines
Nine tasks, and every one lands on an agency rather than on a company:
| Task | Lead | Due |
|---|---|---|
| Prioritize cybersecurity of National Security Systems | CNSS | 2 Jul 2026 |
| Prioritize cybersecurity of DOD information systems | DOD | 2 Jul 2026 |
| Binding Operational Directives; AI defensive tools; extend cyber services to states and critical infrastructure | CISA, OMB, NSC, APNSA, NCD | 2 Jul 2026 |
| Form an AI cybersecurity clearinghouse | Treasury, NCD, NSA, CISA | 2 Jul 2026 |
| Identify federal grant funding for AI vulnerability-detection R&D | OMB, CISA, NCD | 2 Jul 2026 |
| Classified AI benchmarking process + a voluntary framework for covered-frontier-model pre-release access | NSA, NCD, APST, CISA, DOD | 1 Aug 2026 |
| A voluntary framework for AI companies to disclose their models | Treasury, NSA, CISA, NIST, APST, DOD | 1 Aug 2026 |
| Expand US Tech Force cybersecurity hiring pathways | OPM | 1 Aug 2026 |
| Prioritize AI-facilitated cybercrime prosecutions (18 U.S.C. §§1028, 1030, 1343) | DOJ | ongoing |
Funded from existing appropriations, and CRS flags that as unresolved: Treasury "did not specifically request FY2027 funding" for the clearinghouse it is meant to stand up.
The Anthropic suspension is listed as a different instrument
The most consequential paragraph for this wiki is in Considerations for Congress. CRS invites 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 the nonpartisan staff analysis treats the Anthropic suspension as an other approach set against the EO, not as the EO being enforced. This wiki has been reading the suspension and the GPT-5.6 gating as "two enforcement beats" of a standing gate. On CRS's account there was no gate to enforce. See eo-14409 for how both readings are now held.
Two things the corpus did not have
NIST already had pre-release access, by contract, before any of this. Prior agreements between NIST and Anthropic and OpenAI established a framework for NIST "to receive access to major new models from each company prior to and following their public release." Those agreements gave no public transparency and did not cover all frontier labs — but they mean the practice of pre-release government access predates the EO and was arranged commercially rather than by order.
NSPM-11 (5 June 2026), three days after the EO, points the other way: it directs the national security enterprise to work with industry to "make the most advanced frontier models broadly available to national security professionals," while requiring adopted systems be "reliable, robust, steerable, and controllable." Read beside the EO, the pair is slow the model down on the way out, speed it up on the way in.
Legislative context
Two bills would do what the EO does not: the Artificial Intelligence Civil Rights Act of 2025 (H.R. 6356) would mandate pre-deployment evaluation of certain models, and a discussion draft of the Great American Artificial Intelligence Act of 2026 would set transparency, assessment and security-testing frameworks. Neither is law. CRS frames the open question for Congress as whether to keep "the overarching approach of voluntary industry engagement" or establish federal testing requirements.
Related
eo-14409 · us-ai-policy · anthropic-export-ban-2026 · openai-gpt56-access-restriction · caisi · export-controls-on-ai · risk-based-regulation · ai-governance · synthesis