Trump Signs Executive Order Requiring AI Companies to Share Models With US Government
President Donald Trump signed an executive order on Tuesday (local time) requiring AI companies to voluntarily share their models with the federal government **...
President Donald Trump signed an executive order on Tuesday (local time) requiring AI companies to voluntarily share their models with the federal government before public release so the government can assess capabilities — particularly for cybersecurity risks. This is a major shift in US AI policy and sends a clear signal: the era of unchecked frontier model deployment may be ending, and Washington wants a seat at the table before the launch button is pressed.
Background: What is This Executive Order?
The new order comes after weeks of back-and-forth between the White House and tech CEOs. Trump reportedly postponed an earlier signing ceremony because he "didn't like certain aspects of it," according to reporters. The final version is deliberately vague on specifics, according to CNBC, but its core message is clear: the federal government wants early access to the most powerful AI models being developed in the US.
- The order is voluntary — no company is legally forced to comply. But the government will publish evaluations and may designate "trusted partners" who get early access.
- It directs the Department of the Treasury, Department of War (via NSA), Department of Homeland Security (via CISA), and others to create a classified benchmarking process for assessing "advanced cyber capabilities" of AI models.
- Models that cross a certain threat threshold will be labeled "covered frontier model" — opening the door to government oversight and potential restrictions.
This is not the first time Trump has engaged with AI policy, but it marks a more interventionist posture compared to his earlier laissez-faire stance. The tech industry, represented by figures like Elon Musk and David Sacks, had previously opposed a stricter version of this order.
The Core News: What Changed and How It Works
The executive order establishes a voluntary framework where AI developers can engage with the federal government to determine if their model qualifies as a "covered frontier model." If it does, they are asked to provide government access under confidentiality.
Key provisions:
- Within 60 days, the identified agencies must develop a classified benchmarking process to assess AI models' cyber capabilities.
- The Director of the NSA (in consultation with the National Cyber Director, the APST, and others) will determine the threshold for "covered frontier model" designation.
- A voluntary framework will be designed so developers can submit models for evaluation before public release.
- The order explicitly states: "Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement."
Timeline and scope:
| Step | Responsible Party | Deadline |
|---|---|---|
| Develop classified benchmarking process | Treasury, DoW (NSA), DHS (CISA) + others | 60 days from order |
| Define "covered frontier model" threshold | NSA Director + advisory group | Same 60 days |
| Create voluntary engagement framework | Same agencies | 60 days |
| Companies begin submitting models | Voluntary | After framework is ready |
What happens after submission?
The government will assess the model's cybersecurity capabilities (e.g., whether it can be used to launch cyberattacks, find vulnerabilities, or bypass safeguards). If deemed risky, the model may still be released, but the government will have a detailed risk profile — and possibly influence over deployment decisions.
Why This Matters: The Stakes
This executive order sits at the intersection of national security, corporate interests, and global AI competition. Here’s why it’s a big deal:
- Redefines the relationship between Silicon Valley and Washington. Until now, AI companies largely self-regulated. The government is now asking for a pre-release look under the hood — something that could become de facto mandatory if public pressure or elections shift the official stance.
- Creates a new classification system. "Covered frontier model" is a term that will likely be adopted by other countries and international bodies. It’s a marketing and regulatory label that could dictate where and how powerful AI is deployed.
- Impacts IPO timelines. Anthropic just filed confidentially for an IPO; OpenAI is also preparing. Investors may now factor in regulatory uncertainty around model release timing. SpaceX/xAI, meanwhile, is set to go public as soon as next week — potentially less affected because its models are less consumer-focused.
- Sets a precedent for other nations. The EU already has the AI Act with tiered risk categories. China has strict review rules. The US is now carving its own path — voluntary but with hints of coercion.
Comparison of major AI regulatory approaches:
| Dimension | US (Trump EO) | EU AI Act | China |
|---|---|---|---|
| Mandatory? | Voluntary | Mandatory for high-risk | Mandatory for all |
| Pre-release review | Government access before launch | Conformity assessment for high-risk | Government approval required |
| Classification | "Covered frontier model" (cyber-focused) | Risk tiers (minimal, limited, high, unacceptable) | All models subject to content safety laws |
| Enforcement | None (voluntary) | Fines up to 7% of global revenue | Fines, license revocation |
| Scope | Only frontier models with cyber risks | All AI systems | All generative AI |
The US approach is the least restrictive on paper, but the lack of clarity and the government's ability to "select trusted partners" raises concerns about political favoritism and opaque decision-making.
Key Details: Technical Breakdown of the Benchmarking Process
### Who is involved?
- Treasury Secretary – oversees financial cyber risks.
- Secretary of War – explicitly mentioned, likely to counter potential AI-driven military threats.
- NSA Director – leads the classified benchmarking process.
- CISA Director – handles infrastructure vulnerability assessments.
- APST (Assistant to the President for Science and Technology) – coordinates policy.
- NIST Director – sets technical standards.
### What is a "covered frontier model"?
The order leaves the exact threshold to the NSA, but it's likely to be defined by:
- Number of parameters (e.g., >100 billion)
- Capability to autonomously hack or bypass human oversight
- Ability to engineer novel malware or exploit zero-day vulnerabilities
- Potential for dual-use (commercial + military)
### The benchmarking process
- Classified benchmarks will be developed to measure cyber capabilities that cannot be publicly disclosed.
- Developers voluntarily submit their model (or a version of it) to federal labs.
- The government runs the benchmarks and returns a risk assessment.
- Based on the assessment, the government may recommend delaying release, modifying safety features, or sharing findings with other agencies.
This essentially creates a black-box review system — similar to how the government reviews sensitive exports, but applied to software.
Competitive Landscape: Who Wins and Who Loses
### Who benefits?
- National security agencies – get early warning of dangerous AI capabilities.
- Large incumbents (OpenAI, Google DeepMind, Meta) – can afford compliance and may use government partnership as a trust signal.
- AI safety researchers – finally have a mechanism to push for real-world evaluations.
### Who is at risk?
- Smaller startups – don't have resources for government engagement; may be left out of "trusted partner" lists.
- Open-source developers – the order appears aimed at frontier models, not smaller open-source ones, but could be expanded.
- Foreign companies – especially Chinese AI labs like Baidu and ByteDance, which will be excluded from any government trusted-partner framework.
### Industry reaction
- Anthropic (Claude) – likely to cooperate given its safety-first branding.
- OpenAI (Sam Altman) – has already called for regulation, so this aligns with its public stance.
- xAI (Elon Musk) – Musk is a Trump ally and opposed a stricter version; this voluntary order may be a compromise.
- Meta – has open-source LLaMA models; may claim the order doesn't apply to open-source releases.
What This Means for AI-Tool and AI-News Publishers
For publishers covering AI tools, startups, and regulations, this executive order is a goldmine of content angles. Here are concrete story opportunities:
- "Voluntary or just voluntary?" – Write a deep dive on whether voluntary frameworks really work. Use the EO as a case study. Compare to other "voluntary" pledges (like the White House AI commitments in 2023).
- SEO angle: "Covered frontier model" – This term will be searched as people try to understand what it means. Publish a clear explainer targeting keywords like "covered frontier model definition", "AI executive order 2025", "US government AI model evaluation".
- Newsletter tip: Track the 60-day clock. Mark your calendar for when the classified benchmarking process is announced. Write follow-ups on what the process looks like, who gets access, and how companies respond.
- Comparison piece: US vs EU vs China AI regulation – Use the table above as a starting point. Add specific examples of how each region treats models like GPT-5 or Claude 4.
- Listicle: "5 AI companies that will be most affected by the new EO" – Profile Anthropic, OpenAI, xAI, Google, and Meta. Estimate their likelihood of compliance and impact on product roadmaps.
- Opinion: "The end of AI secrecy?" – Argue that this is the first step toward mandatory licensing. Cite expert opinions and historical parallels (export controls on cryptography).
Challenges Ahead and Risks
- Vagueness is a double-edged sword. The order lacks specifics, which gives the administration flexibility — but also creates confusion for developers. A startup might not know if its model qualifies.
- Voluntary ≠ harmless. If the government publishes evaluations or labels certain models as risky, it can effectively pressure companies to comply, creating de facto regulation without law.
- National security vs. innovation. Classifying the benchmarking process means the public won't know what's being measured. Critics worry about over-classification and politicization — e.g., favoring models from politically aligned companies.
- Cybersecurity focus is narrow. The order only addresses cyber capabilities, ignoring other risks like bias, misinformation, or job displacement. This could leave dangerous models unchecked in other domains.
- Global fragmentation. The US, EU, and China now have three distinct regimes. AI developers may need to build different versions of their models for each market, increasing costs and slowing release cycles.
- Legal challenges possible. The order may face lawsuits if companies feel it violates free speech or trade secrets. The voluntary nature makes it less vulnerable, but "trusted partner" selection could be challenged.
Final Thoughts
Trump's executive order is a strategic shift — not a hard regulatory turn, but a foot in the door. By asking for voluntary pre-release access, the government gains de facto oversight without triggering the political backlash of mandatory licensing. The real test will come when a company says no, or when a model fails the secret benchmarking process. That moment will determine whether this order is a footnote or the beginning of a new era in AI governance. For now, developers and publishers alike should watch the 60-day clock and prepare for a more hands-on White House.
FAQ
What exactly did Trump's executive order say?
It asks AI companies to voluntarily provide the federal government with access to their most advanced models before public release so the government can assess cybersecurity risks. It also orders agencies to create a classified benchmarking process and define a "covered frontier model" threshold.
Is this mandatory for AI companies?
No, it's voluntary. The order explicitly states it does not create a mandatory licensing or permitting requirement. But the government can choose to highlight non-compliant companies or only work with "trusted partners," creating soft pressure.
What is a "covered frontier model"?
It's a designation for AI models with advanced cyber capabilities that could pose national security risks. The exact threshold will be determined by the NSA within 60 days. Likely criteria include model size, hacking ability, and potential for dual-use military applications.
How does this affect small AI startups?
Small startups may be at a disadvantage because they lack the resources to engage with the government framework. They could also be excluded from "trusted partner" lists. The classification threshold may also miss smaller but still risky models.
When will the benchmarking process be ready?
The order gives agencies 60 days (by early August 2025) to develop the classified benchmarks and create the voluntary engagement framework. Companies can then start voluntarily submitting models.
What are the biggest risks of this executive order?
Risks include over-classification (secrecy without accountability), political favoritism in selecting trusted partners, a narrow focus on cybersecurity ignoring other harms, and fragmentation of global AI governance. Legal challenges over trade secrets and free speech are also possible.
