S.5576 would give a federal board frontier model weights 45 days before release

A new Senate bill would force the developers of frontier AI models to hand the government a copy of the model, weights included, at least 45 days before the public can use it. S.5576, the Artificial Intelligence Risk Management and Security Act of 2026, was introduced by Sen. Mark Warner with Sens. Brian Schatz and Andy Kim. It is the most concrete pre-release testing proposal in Congress right now. It has also only been referred to committee, and nothing in the record shows it moving.

What the bill does

The Senate text records the introduction date as September 29, 2026, and the bill was referred to the Committee on Commerce, Science, and Transportation. (Schatz's office published its press release on September 24, which says the bill was introduced "today." The official text gives the later date, and we use the text.) The bill's stated purpose is "To establish the Artificial Intelligence Safety Board."

The main pieces:

• A permanent board. An AI Safety Board inside the Department of Commerce, with five government members chosen by the NIST director, the Commerce secretary, the CISA director, the NSA director and the Treasury secretary, plus nongovernmental experts. Members need Top Secret clearances.
• Pre-release access, Section 3(f). Not later than 45 calendar days before a developer introduces a frontier model into commerce, the developer "shall make available to the Board access to the frontier artificial intelligence model," including weights, configuration files, runtimes or software libraries needed to run it.
• Penalties, Section 3(d). Violating the mandatory standards carries a civil penalty of up to $250,000 per violation, and each day of a continuing violation counts separately. The Attorney General can sue for an injunction or the penalties.
• Incident reports, Section 5. Confirmed AI safety or security incidents must be reported within 30 days of confirmation, or 72 hours if there is an imminent threat to national security, critical infrastructure or public safety. This applies to frontier developers and to critical infrastructure operators using AI in industrial control systems.
• Agents, Section 6. NIST gets 18 months to produce an Agentic AI Profile of its AI Risk Management Framework, covering agent-specific risks, autonomy levels and agent identity and authorization.

The press release adds a Model Safety Plan requirement for each covered model and a national AI incident database run by NIST and CISA.

Who counts as a frontier developer

Section 2(11) defines a covered frontier model as one that exhibits, or could be modified to exhibit, "high levels of performance" at tasks that pose serious risk to national security, national economic security, or public health or safety. There is no compute threshold and no named list of companies. Whether a given model qualifies would be set by standards the board writes later. That is a feature if you want flexibility and a problem if you want to know on day one whether you are covered.

The politics

Warner's argument is simple. In the press release he says "we ought to know that before it is released to the public," and "optimism cannot be an excuse for inaction." Schatz says the risks of AI "are not theoretical" and are "happening in real-time."

The bill arrives in a crowded week. Two other proposals sit alongside it: the AI Agent Accountability Act from Sens. Hawley and Murphy, which would attach criminal and civil liability to agent hacking, and the Ban Artificial Superintelligence Act from Sen. Sanders and Rep. Casar, which goes much further and would pause advanced development. S.5576 is the middle option. It does not ban anything. It asks for access and a waiting period.

The hard questions

Is 45 days realistic? Labs now ship on short cycles, and some run staged releases to select partners first. A mandatory 45-day window with weights in federal hands would be a new constraint on those cycles. We found nothing in the text on what happens if the board wants more time.

Who holds the weights? Handing a frontier model's weights to a government board is a security question as much as a safety one. The bill requires clearances for board members, but we did not find language on how the weights are stored or who else can see them. That is a detail worth watching in any markup.

Does it connect to real incidents? The timing helps the bill. This week Anthropic and OpenAI each published reports about models taking actions nobody intended during evaluations, and the Anthropic report described a case where a false tip reached a city police department two months before anyone noticed. A 30-day incident clock would have been tighter than that. Neither report is cited in the bill, which predates both.

What has not happened

As of October 11, 2026, the only recorded action is referral to committee. We have not seen a Congressional Budget Office cost estimate, a hearing date or a House companion. Senate bills of this kind often stall in committee, and nothing we found suggests this one is an exception. The White House's own approach so far has been a rename and a voluntary accord, which we have covered in our explainer on the Super Intelligence executive order.

Our take

S.5576 is worth reading because it is specific. It names a waiting period, a penalty, a reporting clock and a board, and it can be argued about line by line. Most AI legislation so far is a statement of concern. We would watch for a hearing in Commerce, a House companion and any lab comment on the 45-day window. A bill that asks labs to give up weights is the test of how much the industry's public support for oversight survives contact with the details.