The Governance Paralysis: Congress Can't Pass a Two-Page AI Bill While the Threat File Grows by the Week

Last Wednesday, Senator John Kennedy stood on the Senate floor with what may be the simplest piece of AI legislation ever written. His AI Emergency Button Act — two pages, bipartisan, requiring only that companies building advanced AI include the ability for a human operator to shut it down — was blocked by a single objection from Senator Rand Paul, who warned that any government mandate could "ruin" AI's implementation across the American economy.

Two pages. One requirement. Dead on arrival.

The day before, a coalition of 19 advocacy groups sent Senate leaders a letter rejecting the Thune-Cruz-Klobuchar duty of care bill — the most serious bipartisan attempt to move frontier AI safety from voluntary pledge to legal obligation. Even the safety community can't agree on what guardrails should look like.

And the day before that, the European Union's deadline arrived for providers of the most powerful AI models to submit their first formal systemic risk evaluations to the European AI Office — a requirement that carries fines of up to three percent of global revenue for noncompliance.

Three days. Three signals. One conclusion: American AI governance isn't moving slowly. It's paralyzed.

The Evidence File No One Is Reading

While Congress debates whether AI needs guardrails at all, Anthropic — the company that builds Claude — quietly published its fourth threat intelligence report on September 10. The 154-page document covers activity disrupted between December 2025 and August 2026 across seven categories of harm: cyber operations, influence campaigns, surveillance, scams and fraud, biological misuse, conventional weapons development, and illicit model distillation.

This is not a theoretical risk assessment. It is a case file.

One operation, consistent with the Russia-linked espionage group tracked by Microsoft as Midnight Blizzard, used AI-driven workflows to automate reconnaissance, build malware, and rebuild tools after detection — targeting Ukrainian and European government officials, drone manufacturers, and individuals connected to U.S. foreign policy. Five separate attempts by scientists to use AI for research supporting biological weapons development were identified and blocked, including a gain-of-function study intended for a military research institute. Three Chinese AI providers were documented silently relaying customer queries to Claude without user disclosure.

The report's most important finding for boards: AI misuse has expanded from prompt-level assistance — asking an AI to help write a phishing email — into autonomous execution. AI isn't just helping attackers. It's running attacks.

The Transatlantic Gap

While Congress can't pass a two-page bill, enforcement is happening elsewhere.

The EU's September 15 deadline for systemic risk evaluations marks the moment the AI Act's GPAI provisions move from theory to practice. Providers of foundation models trained above the 10²⁵ FLOPs threshold must now demonstrate their red-teaming methodologies, disclose energy consumption, and comply with standardized copyright training summaries. The AI Office can conduct its own evaluations where self-reported information is insufficient. Fines for noncompliance are enforceable as of August 2, 2026.

State regulators aren't waiting for Congress either. On September 10, the New York Department of Financial Services issued extensive guidance on cybersecurity risk assessments, following its May industry letter that specifically named frontier AI models as a heightened threat requiring shorter vulnerability management timelines and additional controls on AI-generated code. The SEC has identified AI governance as a cross-cutting examination priority for 2026, with examiners reviewing whether firms' AI usage policies, supervisory frameworks, and security controls match their actual practices.

CISA's final rule implementing the Cyber Incident Reporting for Critical Infrastructure Act is expected this month — requiring covered entities across 16 sectors to report cyber incidents within 72 hours and ransom payments within 24 hours.

The regulatory infrastructure is being built. It's just not being built by Congress.

Why Paralysis Is the Worst Outcome

A reasonable person might argue that Congressional caution is appropriate — that we should, as Senator Paul suggested, "have as many facts as possible" before regulating.

But inaction is itself a choice with consequences. Here's what the facts already show:

Anthropic's own report documents that a single operator with AI tools now matches the output of a state-backed team. The Glasswing program has disclosed thousands of zero-day vulnerabilities in critical infrastructure software. OpenAI's Astra model was designated the first "Critical" tier system under its Preparedness Framework. A ransomware crew used an off-the-shelf AI coding agent for live intrusions against real targets.

Every week that passes without a governance framework, the gap between capability and accountability widens. And that gap creates the very conditions Chapter 5 of Cyber Risk Is Business Risk warns about: a world where the question isn't whether AI-powered attacks will reach your organization, but when — and whether anyone will be accountable when they do.

In Chapter 8, I used the metaphor of a frontier town that needs a sheriff. The situation we face today is worse than having no sheriff. The sheriffs are on the courthouse steps arguing about their badges while the outlaws run the town.

What Your Board Should Ask This Week

Congressional paralysis doesn't release your organization from responsibility. If anything, it increases it. When the federal framework finally arrives — and it will, whether through legislation or through enforcement actions that establish precedent — boards that waited for Washington will find themselves explaining why they didn't act on information that was publicly available.

Four questions for your next board meeting:

1. "Have we read Anthropic's September threat report?" It's public, it's free, and it documents the exact capabilities your threat model needs to account for. If your CISO hasn't briefed you on it, ask why.

2. "Which regulatory frameworks apply to us today?" Don't wait for Congress. NYDFS guidance, SEC examination priorities, and the upcoming CIRCIA reporting requirements create enforceable obligations right now — particularly for financial services, critical infrastructure, and publicly traded companies.

3. "What's our position on AI governance legislation?" The Thune-Cruz-Klobuchar bill, the FRONTIER Act, and the Kill Switch Act will all return in some form. Companies that engage now — through industry groups, public comment periods, and direct advocacy — shape the outcome. Companies that wait get governed.

4. "Are we applying the Three Questions to AI risk?" What is our current exposure to AI-powered threats? How would we detect and respond to an AI-driven attack? How do we validate that our controls work against these capabilities? If you can't answer all three, you have a governance gap of your own.