US Federal AI Policy Guide: Executive Orders, OMB Memos and Agency Rules

The United States has no federal AI statute, so US federal AI policy lives in executive orders, OMB memoranda and agency actions that change faster than most compliance programmes can track. Since January 2025 the executive branch has revoked the previous AI order, replaced the OMB memos that governed agency AI use and purchasing, added procurement rules for large language models, launched a federal effort to challenge and preempt state AI laws, and in June 2026 issued an order on frontier model cyber capabilities. For security, GRC and public sector teams, and for companies selling AI to US agencies, this guide sets out which instruments are in force as at 24 October 2026, who they bind, what they require, and what they mean for vendors selling AI to government and for security teams running AI agents.

Key takeaways

• Most obligations bind federal agencies, not companies. OMB memos M-25-21, M-25-22 and M-26-04 tell agencies how to govern, buy and contract for AI; private companies feel them through contract clauses and solicitations.
• High-impact AI carries seven minimum practices. Under M-25-21 agencies must test, assess impact, monitor, train operators, keep human oversight, offer appeals and take public feedback for high-impact AI, or stop using it.
• Procurement is the enforcement lever. M-25-22 bars training commercial models on non-public agency data without consent, and M-26-04 makes the Unbiased AI Principles material to eligibility and payment for LLM contracts.
• Preemption is pursued, not yet achieved. EO 14365 created a DOJ AI Litigation Task Force and funding levers against state AI laws, but Congress has not passed a preemption statute and state laws stay in force unless a court or statute says otherwise.
• Identity and logging still decide whether you can prove compliance. ATO requirements, the new OMB logging memo M-26-14 and impact assessments all assume you know which human or non-human identity used which model, with what data.

Scope: who and what is covered

Federal AI policy is a stack of instruments with different legal force. Executive orders direct executive branch officials and state that they create no enforceable rights. OMB memoranda are binding on the agencies they cover. Companies are bound only where a contract, grant condition, regulation or statute applies the requirement to them.

Exclusions that matter: M-25-21 does not apply to national security systems, and the Intelligence Community is encouraged but not required to follow its high-impact practices. M-25-22 excludes the Intelligence Community, national security systems, AI used incidentally by a contractor and common commercial products whose AI is embedded in a substantially non-AI product. M-26-04 excludes national security systems and does not require free, open-source LLMs to meet its contract requirements, although agencies should still perform due diligence.

For the wider picture of state, EU and international rules, see the Global AI Regulation Map.

Key obligations

Agency AI governance (M-25-21)

• Chief AI Officer: retain or designate one within 60 days; in CFO Act agencies at SES, Scientific and Professional, Senior Leader or equivalent level. The CAIO owns the use case inventory, high-impact processes and waivers.
• AI Governance Board: CFO Act agencies within 90 days, chaired at Deputy Secretary level with the CAIO as vice-chair, including cybersecurity, privacy, legal and civil rights representatives.
• AI strategy and compliance plan: within 180 days, posted publicly; compliance plans repeat every two years until 2036.
• Policies: within 270 days, update IT infrastructure, data, cybersecurity (including continuous authorisation for AI) and privacy policies, and develop a generative AI policy.
• Use case inventory: at least annually, submitted to OMB with a public version (DoD and the Intelligence Community excluded).

High-impact AI and the seven minimum practices

AI is high-impact when its output is a principal basis for decisions or actions with a legal, material, binding or significant effect on civil rights, civil liberties or privacy; access to education, housing, insurance, credit, employment and other programmes; access to critical government resources or services; human health and safety; critical infrastructure or public safety; or strategic assets, including sensitive or classified information. Human oversight does not take a use case out of scope. Presumed high-impact categories include biometric identification, biometric access control to IT systems for benefits, benefits eligibility, law enforcement risk assessments, immigration risk assessments and federal employment decisions.

Agencies had 365 days from issuance to document implementation. A use case that cannot meet the practices must be safely discontinued. The CAIO may waive requirements only after a written, system-specific risk assessment; waivers are recertified annually, cannot be delegated and must be reported to OMB within 30 days. Limited pilots certified by the CAIO are exempt. AI systems must still hold an authorisation to operate under OMB Circular A-130.

AI acquisition (M-25-22)

• Data rights: contracts must permanently prohibit using non-public agency data and outputs to further train publicly or commercially available AI without explicit agency consent.
• Lock-in protections: knowledge transfer, data and model portability, government rights to code and models produced under the contract, transparent licensing and pricing.
• Privacy: contract terms for AI handling PII, with the Senior Agency Official for Privacy involved early.
• Performance monitoring: agencies must be able to evaluate performance and risk regularly, independent evaluations should use agency data the vendor cannot see, and contracts must not stop agencies disclosing test results internally.
• High-impact terms: contracts must support the M-25-21 minimum practices, including vendor documentation for the AI impact assessment.

LLM procurement (EO 14319 and M-26-04)

• Two principles: truth-seeking (truthful, historically accurate, acknowledges uncertainty) and ideological neutrality (no intentionally encoded partisan or ideological judgements unless prompted or readily visible to the user).
• Minimum vendor information: an acceptable use policy, model, system or data cards, end-user resources and a mechanism to report outputs that violate the principles. Agencies may ask for more, such as bias evaluation results and enterprise controls, but should avoid demanding model weights where practicable.
• Contract consequences: agencies should make the requirements material to eligibility and payment; under EO 14319 a vendor terminated for non-compliance after a cure period bears decommissioning costs.

Timeline and current status

Status below reflects the primary documents read on 24 October 2026 and, where noted, law firm reporting. Executive orders and memos can be replaced at short notice, so check the Federal Register and the OMB memoranda page before relying on any line.

What has not changed: no federal statute preempts state AI laws as at the date of this guide. State and city laws, such as New York City's hiring law, still apply. See the NYC Automated Employment Decision Tools (AEDT) Law Guide and the California AI Laws Guide.

Enforcement and penalties

There is no federal AI penalty regime equivalent to the EU AI Act. Enforcement works through four channels.

• OMB oversight of agencies: compliance plans, public inventories, waiver reporting and the requirement to discontinue non-compliant high-impact AI.
• Contract remedies for vendors: M-26-04 terms designated material to eligibility and payment can support termination for default; EO 14319 puts decommissioning costs on a vendor terminated for non-compliance. M-25-22 terms on data use and portability are contractual obligations.
• Existing law applied to AI: EO 14365 asked the FTC to explain how the FTC Act's ban on unfair or deceptive practices applies to AI models. EO 14409 directs the Attorney General to prioritise 18 U.S.C. 1028, 1030 and 1343 against people who use AI, including AI agents, to access computers or data unlawfully.
• Federal pressure on states: litigation by the AI Litigation Task Force, possible loss of BEAD non-deployment funds for states with laws Commerce identifies as onerous, and conditions on discretionary grants. These target states, not companies, but a successful challenge changes which state obligations companies face.

Decision guidance: a company that does not sell to government and is not a frontier developer has few direct federal AI obligations, and its AI compliance effort should follow state laws, sector regulators and the EU AI Act where relevant. A federal contractor should treat M-25-22 and M-26-04 terms as the baseline for every AI offering, because they appear in solicitations. A large frontier developer should assess the voluntary EO 14409 framework and its 30-day pre-release access arrangement with legal counsel.

What it means for AI agents and non-human identities

None of these instruments says "non-human identity", but each assumes you can attribute AI activity to an identity and evidence it.

• Authorisation to operate: M-25-21 and M-25-22 keep the A-130 ATO requirement for AI systems. In practice the assessor will ask how the agent or model service authenticates to agency data, which maps to NIST SP 800-53 IA-9 (Service Identification and Authentication), AC-2 (Account Management) and AC-6 (Least Privilege). See the FedRAMP Guide.
• Logging identity: M-26-14 (22 May 2026) rescinded M-21-31 and requires logs that show which identity performed operations in applications and systems, searchable for at least 6 months and retrievable for at least 1 year at baseline. Agent service accounts, API keys and OAuth clients need distinct identities, or you cannot meet that test. See the Log Management and Retention Guide.
• Use case inventories: an AI use case inventory is only complete if it includes agents running under service identities, not only user-facing tools. Pair it with an NHI inventory. See the Identity Governance for AI Agents Guide.
• Human oversight and fail-safes: for high-impact AI the oversight practice implies the ability to suspend an agent quickly, which in turn means revocable credentials and a named owner. See the AI Agent Authorisation Guide.
• Criminal enforcement of agent misuse: EO 14409 explicitly names AI agents used to access data unlawfully. Incidents such as the OpenAI agents and US government websites case, where training agents found other people's Census Data API keys in public GitHub repositories and used them, show why developer API keys need secret scanning, monitoring and rotation.

Compliance steps

1. Classify your exposure. Decide whether you are an agency, a contractor, a frontier developer or none of these; the obligations differ completely.
2. Inventory AI use cases and agents. Record purpose, model, data, owner and the identity each agent runs as. Flag anything matching the presumed high-impact categories.
3. Write the high-impact determination. For presumed high-impact cases judged not high-impact, M-25-21 requires a written determination to the CAIO; vendors should supply the facts that support it.
4. Build the evidence pack. Testing results, impact assessment inputs, monitoring metrics, the M-26-04 documentation set and a data-use statement covering the M-25-22 training prohibition. See the AI Impact Assessment Template.
5. Check contract terms before bidding. Confirm data training restrictions, portability, testing disclosure and decommissioning cost clauses; price them in.
6. Instrument identity and logging. Give every agent and model integration its own identity, log calls with that identity, and meet M-26-14 retention where you operate agency systems.
7. Pilot and roll back safely. Run new high-impact use cases as CAIO-certified pilots limited in scale and duration, with a documented discontinuation plan if minimum practices cannot be met.
8. Track change monthly. Watch the Federal Register, the OMB memoranda page and DOJ for new orders, memos and Task Force actions, and record the date of each check.

Metrics that show it is working: percentage of AI use cases with a documented high-impact determination; percentage of high-impact cases with all seven practices evidenced; number of active CAIO waivers and days to recertification; percentage of AI agents with a unique identity and named owner; contracts with M-25-22 and M-26-04 terms as a share of AI contracts; days since the last regulatory status check.

Common failure patterns

• Working from revoked documents: teams still cite EO 14110, M-24-10 or M-24-18 in policies and bids, all of which have been revoked or rescinded.
• Assuming preemption has happened: pausing state compliance because of EO 14365. An executive order cannot itself repeal state law; only courts and Congress can remove it.
• Inventories that miss agents: listing chat tools but not background agents running on service accounts, which then fall outside impact assessment and logging.
• No documentation set for LLM bids: vendors unable to produce an acceptable use policy, model cards and a reporting mechanism when the solicitation asks, which M-26-04 makes material.
• Shared credentials behind AI services: one API key for many agents makes per-identity logging under M-26-14 and incident scoping impossible, and a key leaked once is reused by anyone who finds it, human or agent.

Practitioner checklist

• Remove references to EO 14110, M-24-10 and M-24-18 from policies, templates and bids.
• Map each AI use case against the M-25-21 high-impact definition and presumed categories.
• Hold evidence for all seven minimum practices for every high-impact use case.
• Prepare the M-26-04 documentation set (acceptable use policy, model or system cards, end-user resources, reporting mechanism) for LLM offerings.
• Confirm contracts prohibit training on non-public agency data without consent and support portability.
• Give every AI agent and model integration a unique, owned, revocable identity.
• Log model and agent activity per identity with retention that meets M-26-14 where it applies.
• Keep state law compliance running until a court or statute says otherwise.
• Review the Federal Register, OMB memoranda and DOJ Task Force activity monthly and record the check.

Standards and references

• OMB: M-25-21 Accelerating Federal Use of AI through Innovation, Governance, and Public Trust
• OMB: M-25-22 Driving Efficient Acquisition of Artificial Intelligence in Government
• OMB: M-26-04 Increasing Public Trust in Artificial Intelligence Through Unbiased AI Principles
• Federal Register: Executive Order 14365 Ensuring a National Policy Framework for Artificial Intelligence
• Federal Register: Executive Order 14409 Promoting Advanced Artificial Intelligence Innovation and Security
• Federal Register: Executive Order 14319 Preventing Woke AI in the Federal Government
• US Department of Justice: Artificial Intelligence Litigation Task Force memorandum
• OMB: M-26-14 Ensuring Effective and Efficient Agency Logging and Network Visibility

Related NHI Mgmt Group resources: Global AI Regulation Map · Public Sector Identity Security Guide · FedRAMP Guide · AI Agent Authorisation Guide. This guide summarises US federal AI executive orders, OMB memoranda and related actions as at 24 October 2026 and is not legal advice.