The Trump administration is studying an executive order that would formalize pre-deployment safety reviews for frontier AI models — explicitly modeled on the way the Food and Drug Administration evaluates drugs. National Economic Council director Kevin Hassett framed the Trump AI executive order as “a clear road map” for how new frontier models can be “proven safe” before release. Reporting indicates the order would stop short of mandatory government approval but would partner agencies with AI companies on cybersecurity and biosecurity testing, building on the Center for AI Standards and Innovation (CAISI) pre-deployment agreements already in place with Anthropic, OpenAI, Google DeepMind, Microsoft, and xAI.
What’s actually new
The Trump AI executive order under consideration represents a notable shift for an administration that began with a deregulatory posture toward AI. Through early 2026, the White House signaled “hands-off” preference and rolled back several Biden-era AI executive orders. The current direction — even with the explicit FDA-style framing — reflects rising concern over the cybersecurity and biosecurity implications of frontier models. Anthropic’s Claude Mythos and OpenAI’s GPT-5.5 both crossed thresholds on offensive cyber capability through 2026 that drew White House attention.
The reported structure of the order. Mandatory pre-deployment partnership with US agencies on cybersecurity and biosecurity testing. Not mandatory approval — the administration appears to want a process closer to FDA’s “go through testing, then ship” model than a “wait for government sign-off” gate. Coordination via CAISI, which has already conducted 40+ evaluations of frontier models including unreleased variants. Focus areas: cyber-attack capability, biological weapon-design assistance, chemical weapon synthesis assistance.
The FDA comparison is deliberate. The Hassett framing emphasizes that drugs go through formal testing before going to market, and AI models — given their stakes — should follow similar discipline. The political appeal is real: the FDA-style framing positions oversight as standard-issue safety practice rather than novel regulation. Critics note the FDA does require approval, not just testing, while the AI proposal seemingly does not.
The CAISI background. The Center for AI Standards and Innovation sits within NIST and conducts pre-deployment evaluations of frontier AI under agreements with the major labs. Anthropic and OpenAI signed first; Google DeepMind, Microsoft, and xAI joined in early May 2026. The agreements are voluntary; the proposed executive order would formalize the framework and likely expand scope.
Why it matters
- It signals a posture shift on AI regulation. The 2025-early-2026 narrative was “Trump administration favors AI deregulation.” The Trump AI executive order discussions complicate that narrative.
- It sets the formal US framework for frontier AI safety review. Voluntary CAISI agreements become institutional. The EU AI Act has had a more aggressive framework; the US would have its own.
- It positions cybersecurity and biosecurity as the central concerns. Other AI policy issues (bias, employment displacement, copyright) get less explicit treatment.
- It complements rather than replaces state AI laws. Earlier Trump executive orders explicitly preempted state AI regulation; this order is reportedly federal-level safety review, not state-law preemption.
- It maintains industry control over development. The “no mandatory approval” structure means companies remain decision-makers; government participates in safety review but doesn’t gate releases.
- It impacts the competitive landscape internationally. US AI labs facing CAISI review will need to integrate this into release timing; Chinese labs face no such requirement and may iterate faster.
How to use it today
The Trump AI executive order is in study phase; nothing is signed as of mid-May 2026. Here’s how to engage practically.
- Read the source material. The reporting comes from Bloomberg, NYT, Washington Post, Federal News Network and others. Read multiple sources for the full picture.
# Source coverage Bloomberg: "US Prepares AI Security Order That Omits Mandatory Model Tests" NYT: original reporting on the considered order Federal News Network: "WH 'studying' AI security executive order" CAISI: nist.gov/centers/center-ai-standards-and-innovation - Track the formal announcement. When the executive order is signed (timing TBD), the text will be at whitehouse.gov/presidential-actions and the Federal Register.
# Where formal announcements appear whitehouse.gov/presidential-actions/ — primary source federalregister.gov — official publication NIST press releases — CAISI updates - If you’re at a frontier AI lab, prepare for CAISI engagement. The current CAISI process involves pre-deployment evaluations. If the executive order formalizes this, expect:
# Likely CAISI evaluation scope - Cybersecurity capability assessment - Biosecurity / dual-use chemistry assessment - Specific tested capabilities (offensive security, biological design) - Time investment: weeks to months per model - Confidentiality structure for unreleased models - Reporting back to the lab on findings - If you’re an enterprise AI buyer, factor regulatory posture into vendor evaluation. Vendors who engage cooperatively with CAISI signal commitment to safety review.
- If you’re a policy professional, identify your stakeholders. Different parts of the AI ecosystem will have different stakes — academic researchers, open-source advocates, enterprise IT, civil society — each with positions on the framework.
- If you’re a developer building AI products, the order likely doesn’t directly affect you. Pre-deployment review applies to frontier-model labs, not downstream applications. Build as before; track your model providers’ release timing changes.
- If you’re an AI safety researcher, the order may create funding and engagement opportunities. CAISI capacity expansion will need talent. Government and contractor positions may proliferate.
- Watch state-level interactions. The earlier Trump executive orders preempted some state AI regulation. How this order interacts with state laws (California’s AB 1018, Colorado’s AI Act, etc.) will matter.
# Tracking state AI laws NCSL: tracks state legislation IAPP: AI legislation tracker Specific state government sites
How it compares
The Trump AI executive order under consideration sits in a global regulatory landscape worth understanding.
| Framework | Scope | Approach | Status |
|---|---|---|---|
| Trump AI executive order (proposed) | US frontier AI models, cybersecurity + biosecurity | CAISI partnership, no mandatory approval | Under consideration, May 2026 |
| EU AI Act | AI broadly, with risk-tiered obligations | Mandatory compliance, conformity assessments for high-risk | Phased enforcement through 2025-2027 |
| China AI regulations | Generative AI deployment, content rules | Mandatory registration, content compliance | Active |
| UK AI white paper approach | Sector-specific regulator-led | Existing regulators apply principles | Implementation phase |
| Singapore Model AI Governance Framework | Voluntary best-practices | Soft governance, sandbox programs | Updated 2024-2026 |
| California state laws (AB 1018, etc.) | Specific use cases within California | Mandatory for in-scope deployments | Varies by law |
What distinguishes the considered Trump AI executive order: focus on national-security-adjacent capabilities (cybersecurity, biosecurity, chemistry) rather than broad civil-rights or consumer-protection emphasis; no mandatory approval gate; building on existing voluntary CAISI agreements; FDA-style framing. The risks: implementation details matter substantially; “no mandatory approval” can become de facto required compliance; coordination complexity across many AI labs and use cases.
What’s next
Signals to watch through summer 2026. The actual signing: timing is unclear. Could be weeks; could slip into fall. Watch presidential-actions feed at whitehouse.gov. The specific text: drafts may differ from reporting. Coverage focus areas, agency designations, timeline requirements, and enforcement mechanisms all matter. Industry response: the major labs have shown cooperation via CAISI agreements. Whether they support formalization or push back will signal real industry posture. State law preemption: this order’s relationship to California, Colorado, and other state AI laws will shape what compliance looks like for AI products operating across states.
The broader implications. The Trump AI executive order discussions complete the picture of US federal AI governance. Combined with the earlier executive orders preempting state law and the CAISI agreements, a recognizable federal framework emerges: deregulatory on civil-rights and consumer-protection dimensions; security-focused on cyber and bio dimensions; state-law preempted on conflicts with federal direction. Whether this framework holds, evolves, or reverses in subsequent administrations matters for long-term AI deployment patterns.
For the AI industry, the practical impact in the next 6-12 months depends on how the order, once signed, defines scope and process. If pre-deployment review adds 2-4 weeks to flagship-model releases without blocking them, industry can absorb that. If it extends to 6+ months or creates approval bottlenecks, the calculus changes. Watch the implementation details closely.
Frequently Asked Questions
Has the Trump AI executive order been signed?
Not as of mid-May 2026. Reporting from May 4-8, 2026 indicates the administration is “studying” the order. Signing timing is unclear. Watch whitehouse.gov/presidential-actions for official announcements.
How does this differ from the Biden-era AI executive order?
President Trump revoked or substantially modified Biden’s AI executive orders early in his term. The considered new order would replace some of that framework with a different emphasis — cybersecurity and biosecurity rather than broader civil-rights and consumer-protection focus, and partnership rather than mandatory disclosure.
Does this require government approval before releasing AI models?
Reporting indicates no. The structure is FDA-evaluation-style — go through testing and prove safety, but releases don’t require government sign-off. Implementation details may shift this.
What’s CAISI’s role in all of this?
The Center for AI Standards and Innovation within NIST conducts pre-deployment evaluations of frontier AI under voluntary agreements with major labs. CAISI would likely be the operational vehicle for the proposed executive order’s review process. CAISI has already conducted 40+ evaluations including on unreleased models.
How does this affect open-source AI models?
Unclear from current reporting. Pre-deployment review fits the closed-frontier-model lab pattern (Anthropic, OpenAI, etc.) more than open-weights releases (Meta Llama, Mistral). The order’s text will need to address how open releases factor in.
What about international AI companies operating in the US?
The order’s scope on non-US companies isn’t clear from reporting. US-deployed AI typically falls under US frameworks regardless of company origin. Companies like Mistral (France) or international branches of major labs would need to evaluate per the order’s specific text.