US AI Executive Order 14110: Red-Teaming & Dual-Use Compute Thresholds
The Federal AI Mandate: Executive Order 14110 Overview
On October 30, 2023, President Joe Biden signed Executive Order 14110, titled Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence (*White House Executive Actions: EO 14110*, Oct 2023).
Spanning 110 pages across eight distinct policy pillars, EO 14110 represents the most sweeping federal intervention into the domestic computer science and software industry in United States history. While federal legislation in Congress remained stalled, the White House utilized existing executive authorities to establish enforceable technical standards for developers training frontier artificial intelligence systems.
To establish binding legal authority over private technology companies without new statutory legislation from Capitol Hill, the Executive Order strategically invoked the Defense Production Act (DPA) of 1950 (50 U.S.C. Β§ 4501 et seq.)βa Cold War-era statute granting the President expansive authority to compel private enterprise to disclose industrial capacity and technical data essential to national defense.
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β US Executive Order 14110 Regulatory Architecture β
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β β
β Invoking Defense Production Act (DPA) Authorities β
β Administered by US Department of Commerce / BIS β
β β β
β ββββββββββββββββββββββββββ΄βββββββββββββββββββββββββ β
β βΌ βΌ β
β ββββββββββββββββββββββββββββββββ ββββββββββββββββββββββββββββββ β
β β Frontier Model Developers β β Cloud Infrastructure (IaaS)β β
β ββββββββββββββββββββββββββββββββ€ ββββββββββββββββββββββββββββββ€ β
β β Mandatory Federal Reporting: β β Mandatory Foreign KYC: β β
β β - Compute > 10^26 FLOPs β β - Track foreign tenants β β
β β - Bio/Chem AI > 10^23 FLOPs β β leasing large clusters β β
β β - Mandatory Red-Team Results β β - Report foreign training β β
β β - NIST AISI Safety Audits β β runs exceeding threshold β β
β ββββββββββββββββββββββββββββββββ ββββββββββββββββββββββββββββββ β
β β β
β βΌ β
β ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ β
β β National Institute of Standards and Technology (NIST) β β
β β - US AI Safety Institute (AISI) Technical Benchmarking β β
β ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ β
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
---
Dual-Use Compute Thresholds: $10^{26}$ and $10^{23}$ FLOPs
The core mechanism of EO 14110 centers on defining what constitutes a "dual-use foundation model"βa model with broad capabilities that could be weaponized for cyber offensive operations, biological warfare, or national infrastructure sabotage.
The Executive Order established strict mathematical thresholds triggering mandatory federal disclosure: 1. The General Frontier Compute Threshold: Any model trained using a quantity of computing power greater than $10^{26}$ integer or floating-point operations (FLOPs) is legally classified as a dual-use foundation model. 2. The Biological / Chemical Threshold: Recognizing the asymmetric risk of automated pathogen synthesis, the threshold drops by three orders of magnitude to $10^{23}$ FLOPs for models trained primarily on biological, chemical, or radiological sequence data. 3. Mandatory Reporting Obligations: Companies meeting these thresholds must submit quarterly reports to the Department of Commerce detailing: - Complete physical security protocols and weight encryption standards. - The ownership, geographic location, and hardware composition of the training cluster. - Comprehensive results of internal and external red-teaming evaluations before commercial release.
---
Cloud Infrastructure (IaaS) Reporting and "Know Your Customer" Rules
Crucially, EO 14110 does not solely regulate model developers; it imposes sweeping compliance mandates on commercial cloud Infrastructure-as-a-Service (IaaS) hyperscalers (Amazon Web Services, Microsoft Azure, Google Cloud, Oracle Cloud): - Foreign Customer Reporting (KYC): Cloud providers are legally required to verify the identity of foreign persons who contract for or open accounts to train models exceeding the $10^{26}$ FLOP threshold. - Notification Mandates: When a foreign entity leases a compute slice capable of executing a covered training run, the IaaS provider must immediately notify the Department of Commerce, detailing the customer's legal identity, IP ranges, and leased hardware capacity. - Export Control Synergy: This provision works in tandem with the Department of Commerce Bureau of Industry and Security (BIS) export controls, preventing foreign adversaries from circumventing physical GPU export bans by renting US-based cloud clusters.
---
The NIST AI Safety Institute (AISI) and Red-Teaming Standards
To translate executive policy into empirical measurement, EO 14110 directed the National Institute of Standards and Technology (NIST) to establish the United States AI Safety Institute (AISI): - Standardized Red-Teaming: AISI was tasked with authoring rigorous, reproducible red-teaming benchmarks to evaluate model risks across four critical threat vectors: 1. *Chemical, Biological, Radiological, and Nuclear (CBRN)* weapon synthesis assistance. 2. *Automated Cyber Offense*, including autonomous zero-day discovery, exploit generation, and firewall evasion. 3. *Autonomous Self-Replication* and model exfiltration capabilities. 4. *Deception and Influence Operations*. - Pre-Deployment Safety Agreements: In August 2024, the US AI Safety Institute signed historic formal Memoranda of Understanding with Anthropic and OpenAI. Under these agreements, the federal government receives formal access to major new foundation models *prior to and following* public deployment to conduct independent safety evaluations.
---
Comparing EO 14110, California SB 1047, and the EU AI Act
The global regulatory landscape for artificial intelligence has divided into three distinct operational philosophies:
| Regulatory Framework | United States (EO 14110) | California (SB 1047 - Vetoed) | European Union (EU AI Act) | | :--- | :--- | :--- | :--- | | Legal Mechanism | Executive Order / DPA Cold War Authority | Proposed State Statutory Legislation | Binding Transnational Parliamentary Law | | Risk Classification | Compute Threshold ($10^{26}$ FLOPs) | Compute ($10^{26}$) + Cost Threshold ($100M) | Risk Tiers (Unacceptable, High, GPAI, Minimal) | | Primary Focus | National security, defense, CBRN threats | Catastrophic tort liability & kill switch | Human rights, data privacy & market fairness | | Enforcement Body | Department of Commerce & NIST AISI | California Attorney General | European AI Office & National Authorities | | Corporate Penalty | DPA civil/criminal enforcement | Injunctions & up to 10% model cost | Fines up to β¬35M or 7% global turnover |
---
The Constitutional and Political Fragility of Executive Orders
While EO 14110 established a functioning federal oversight apparatus, legal scholars and industry analysts recognize its inherent institutional vulnerability: 1. The Limits of the Defense Production Act: Critics argue that stretching the 1950 DPAβoriginally drafted to secure domestic steel production during the Korean Warβto compel commercial software developers to hand over proprietary neural network weights invites severe legal challenges under the Administrative Procedure Act (APA) and the Major Questions Doctrine. 2. Vulnerability to Executive Reversal: Because EO 14110 was enacted by presidential decree rather than congressional statute, any future administration can unilaterally rescind, modify, or halt enforcement with a single signature. 3. The Unfinished Legislative Mandate: Lasting federal AI governance in the United States will ultimately require statutory legislation from Congress establishing clear intellectual property boundaries, developer liability frameworks, and permanent regulatory bodies.
Despite these political uncertainties, Executive Order 14110 succeeded in embedding institutional safety evaluations, cybersecurity mandates, and federal reporting directly into the frontier development workflows of labs training systems like Claude and ChatGPT.