Sens. Mark Warner (D-VA), Brian Schatz (D-HI), and Andy Kim (D-NJ) · 2026
Artificial Intelligence Risk Management and Security Act of 2026 (S. 5576)
AI Risk Management and Security Act
Introduced September 29, 2026 and referred to the Commerce Committee, S. 5576 was put up for passage by unanimous consent the same day, and Sen. Ted Cruz objected, saying Congress must not legislate on AI 'hastily or in a closed manner.' The bill creates a permanent Artificial Intelligence Safety Board within Commerce. Its government members come from NIST, Commerce, CISA, the NSA, and Treasury, and they sit alongside outside technical experts, some unaffiliated with AI developers and some affiliated, plus evaluation and consumer-protection specialists, all cleared to the Top Secret/SCI level and briefed by the intelligence community. Within 90 days the Board must propose standards: a process for deciding which capabilities pose a serious risk to national security, economic security, or public safety; risk metrics that cover unreleased checkpoints; formats for model cards; cybersecurity, insider-risk, and resourcing recommendations; and security controls for evaluation environments, including real-time monitoring and prompt termination of any evaluation in which a model escapes its controls. The Secretary must adopt them by rule within 30 days, and compliance is mandatory, enforced by civil penalties of up to $250,000 per violation per day and injunctions sought by the Attorney General. Every frontier developer must give the Board access to a model, including its weights, at least 45 days before release, and must file a Model Safety Plan for each model it creates, substantially modifies, or uses to train or evaluate others. Developers and providers of frontier models, and critical-infrastructure operators that use AI in industrial controls, must report confirmed safety or security incidents within 30 days, or 72 hours if they threaten national security, critical infrastructure, or public safety, into a NIST–CISA database that also tracks recurring flaws across models. NIST must also develop an agentic AI profile of its risk management framework and a common template for documenting AI agents.
Key Provisions
Permanent Artificial Intelligence Safety Board in Commerce: members selected by NIST, Commerce, CISA, the NSA, and Treasury, plus nongovernmental experts both unaffiliated with and affiliated with developers; Top Secret/SCI clearances; public conflict-of-interest disclosures; staffed by CAISI
Board standards due within 90 days and adopted by the Secretary by rule within 30 days: risk-capability evaluations, metrics covering unreleased checkpoints, model card formats, cybersecurity and insider-risk controls, and secure evaluation environments with real-time monitoring and termination conditions
Mandatory compliance with adopted standards; civil penalties up to $250,000 per violation, each day a separate violation; Attorney General suits for injunctions
Developers must give the Board access to each frontier model, including weights, configuration files, and runtimes, at least 45 days before release
Model Safety Plan for every model a developer creates, substantially modifies, or uses to train or evaluate other models, with a capabilities and risk assessment, mitigations, and a responsible officer, filed with Commerce
Mandatory incident reports within 30 days, or 72 hours for imminent threats, from frontier developers and providers and from critical-infrastructure operators using AI in operational technology, covering incidents during development, testing, and red-teaming as well as deployment
Public NIST–CISA database of AI safety and security incidents with a catalog of recurring AI flaws across models and providers
Secure test beds at the NSA and the national laboratories, available on a cost-recovery basis for pre-deployment testing
International agreements on joint testing, regulatory reciprocity, and incident information sharing
NIST agentic AI profile and common assurance template for AI agents within 18 months
Regulatory Philosophy
Government sees the model before the public does. Where most frontier proposals rely on developers publishing their own assessments, S. 5576 gives a federal board the model itself, weights included, 45 days before release, and makes the board's standards binding. It is built from the national-security side of government — NSA, CISA, and cleared experts with access to classified threat intelligence — and treats frontier AI as a security problem to be managed by people who can see both the models and the threats. It stops short of a release gate: the board evaluates and sets standards, but nothing in the text authorizes it to block a release once the 45 days have passed.
Developers of frontier models defined by dangerous capability rather than compute, plus providers of those models for incident reporting
Why it lands there
Standards, pre-release access, and safety plans fall on frontier developers. Incident reporting extends to anyone who provides a frontier model, which pulls in the hyperscalers that host them, and to critical-infrastructure operators using AI in industrial controls — a category outside the four industry layers entirely. App companies built on frontier models are untouched unless they provide the model themselves.
September 29: a voluntary accord and a mandatory board
The White House's superintelligence order and its accord with frontier AI executives date from September 29, 2026, the day Warner, Schatz, and Kim introduced S. 5576 and asked the Senate to pass it by unanimous consent. The accord asks companies to monitor their own models and choose their own auditors, binding no one. The bill would hand a federal board the models themselves, weights included, 45 days before release, and make its security and evaluation standards enforceable at $250,000 a day. Cruz's objection kept the voluntary approach as the only one in force.
+The 45-day pre-release access to weights is the most direct government look at unreleased models of any introduced bill on the map
+Standards adopted by the board are mandatory, with daily penalties, rather than voluntary guidance
+Incident reporting reaches development, training, testing, and red-teaming, not only deployment, and extends to critical-infrastructure operators using AI in industrial controls
+Tracking recurring flaws across models and providers treats AI failures the way aviation and cybersecurity treat systemic defects
+Secure-evaluation-environment standards, including termination conditions, respond directly to the risk of a model escaping its test environment
Weaknesses
From the perspective of political opposition
−The frontier model definition turns on whether a model 'exhibits or could be modified to exhibit' high performance at dangerous tasks, with no compute or revenue threshold, so coverage is uncertain until the board defines it
−Access 45 days before release is not a gate: the bill gives the board no power to delay or block a model it finds dangerous
−Ninety days for the board to produce a full suite of technical standards, and thirty for the Secretary to adopt them by rule, is an implausible timeline
−Board members include experts affiliated with AI developers, which invites capture of a body whose standards bind those developers
−Silent on preemption, so it would layer onto California and New York frontier laws rather than settle the federal-state question
−Introduced without Republican cosponsors and blocked on its first unanimous-consent request