On 24 September 2026, Democratic Sens. Brian Schatz (Hawaii) and Mark R. Warner (Virginia), vice chair of the Senate Select Committee on Intelligence, introduced the Artificial Intelligence Risk Management and Security Act of 2026 and announced Senate-floor remarks to push it forward. Later coverage also placed Sen. Andy Kim in the joint legislative pressure.
At the core is a permanent Artificial Intelligence Safety Board inside the Department of Commerce, with NIST, Commerce, CISA, NSA, Treasury, and independent technical experts. Its mandate: evaluate emerging risks and set technical safety and cybersecurity standards for frontier systems.
The industry’s most contested duty: frontier-model developers would give the Board access to their models at least 45 days before public release, including weights, configuration files, runtimes, and libraries needed to run the model. Not a marketing summary—material sufficient for pre-deployment evaluation.
They would also require Model Safety Plans: capabilities and risks, concrete mitigations, and the corporate officer responsible for the plan. Failure to meet applicable standards would face civil penalties up to $250,000 per violation per day, per the senators’ release.
The bill creates a secure reporting process and a national AI incident database at NIST, coordinated with CISA, to track failures, near misses, and recurring patterns. Serious incidents would generally be reported within 30 days, and within 72 hours when there is an imminent threat to national security, critical infrastructure, or public safety.
Secure federal test environments could draw on NSA and Department of Energy national-lab resources. The text also calls for standards specific to autonomous agents—identity, authentication, authorization, data access, and autonomy levels—and standardized documentation of intended uses, authority limits, and independent-evaluation results.
Warner and Schatz frame urgency around warnings from the labs themselves—OpenAI, Anthropic, Google DeepMind, Meta, Microsoft—about frontier models’ cyber-offensive capabilities, plus “clear failures” in security practices in operating and deploying models. “The risks of AI are not theoretical—they are happening in real-time,” Schatz said.
Neural Pulse notes the timing: introduction lands in the same news window as agent escapes, Australian Medicare, and training pauses. That does not make the bill law; it does raise the odds of hearings and cyber-focused bipartisan amendments.
Industry will object to 45-day weight access: IP leakage, competitive delay, and risk the Board becomes a political bottleneck. The authors reply that if a model can find vulnerabilities in a bank or power grid, better to know before the public does.
Scope is limited, per the release, to advanced systems that exhibit—or could be modified to exhibit—high performance on tasks posing serious risks to national security, economic security, or public health and safety. The exact “frontier” definition will be a regulatory battlefield.
Compared with FTC Chair Ferguson’s line (existing tools, do not anthropomorphize), the Schatz-Warner bill is maximalist on new institutions and pre-release access. Both, however, refuse to leave governance solely to the lab.
For startups below the frontier threshold, the risk is contagion: enterprise buyers may demand “Safety Board–like” compliance even if the statute does not apply. Product counsel should watch how the perimeter is defined.
Neural Pulse treats the text as a 24 September proposal with announced teeth—45 days, weights, daily fines—not as a regime already in force. Committee process and negotiation with the Republican majority will decide what survives.
Close: in a month saturated with agents outside sandboxes, Warner and Schatz bet Congress will treat model cybersecurity with the seriousness of critical-infrastructure security—or at least force a debate with numbers and deadlines.
