---
title: "AI Regulation Targets Models. Chips and Agents Decide More"
slug: "ai-regulation-targets-models-chips-and-agents-decide-more"
author: "Miguel Amigot"
date: "2026-11-19 12:00:00"
category: "Premium"
sector: "cross"
topics: "AI regulation, AI governance, chip export controls, Nvidia H200 China, frontier AI licensing, AI liability, AI LEAD Act, Artificial Intelligence Risk Evaluation Act, EU AI Act, SB 53, FBI ARTEX, AI agent security, deployment governance"
summary: "Most AI regulation targets the model with licensing, pre-deployment evaluation and liability, a playbook built for hazards that sit in one place. Chip export licensing decides who can build frontier AI, and the deployment layer decides what agents do: the FBI's ARTEX alert describes an attack agent that brings no model of its own."
banner: ""
thumbnail: ""
linkedin: |
  Most of the AI regulation debate is about the model: license it, evaluate it before release, make its developer strictly liable.

  It is a playbook built for hazards that sit in one place. Two other layers are deciding more.

  Upstream, chips. In January 2026 the Commerce Department moved exports of Nvidia's H200, AMD's MI325X and similar chips to China to case-by-case license review, with conditions including independent U.S. testing. One agency's licence decision moves more frontier capability than any model bill now in Congress.

  Downstream, deployment. On November 13 the FBI described ARTEX, an open-source attack agent that "does not include its own" model. It plugs into Claude, GPT, DeepSeek, GLM, Grok or Kimi, and some intrusions reached privileged access in under three hours.

  A rule aimed at one model cannot follow an agent that swaps models. The layer an organization can govern today is its own deployment: which agents exist, what tools each may call, and what they did.

  With ibl.ai you own all the code and the data, so that layer runs on your infrastructure, on any model, with no per-seat pricing.

  #iblai #AgenticAI #EnterpriseAI #AIGovernance #AIRegulation #ExportControls
faqs:
  - question: "Is AI regulated through the model or through how it is used?"
    answer: "Mostly through the model so far. The EU AI Act's general-purpose model obligations have applied since August 2, 2025, and California's SB 53 was signed on September 29, 2025. U.S. federal proposals for model licensing, pre-deployment evaluation and strict liability remain bills. Chip export licensing and deployment controls do more of the practical work."
  - question: "Does the FBI ARTEX alert show that model regulation does not work?"
    answer: "It shows its limit. The FBI says ARTEX includes no model of its own and connects to frontier and open models alike, with Chinese models most prevalent in the instances analysed. A rule on one model's developer does not reach a tool that can switch to another model."
  - question: "What can an organization govern while AI laws are still being written?"
    answer: "Its own deployment: which agents exist, which tools each may call, what is screened before and after the model answers, and a record of what each agent did. That layer is governable today, and it is easier to govern when the organization runs it on infrastructure it controls."
---

## The Short Answer

**Most AI regulation debate targets the model, but two other layers decide more: export licensing of chips decides who can build frontier AI, and the deployment layer decides what agents actually do. The FBI's ARTEX alert shows an attack agent that brings no model of its own. On ibl.ai, you own all the code and the data, so the deployment layer is one you govern yourself.**

This is not an argument that model rules are pointless. It is an argument about where the leverage sits in November 2026, and which layer an organization can act on before any of the pending bills become law.

## What does model-centric AI regulation actually propose?

Three tools, each aimed at the developer of a powerful model. The first is a licence. In May 2023, OpenAI's Sam Altman [told a Senate Judiciary subcommittee](https://www.judiciary.senate.gov/download/2023-05-16-testimony-altman) that the government should consider "licensing or registration requirements" for models above a capability threshold.

That idea became the first pillar of the [Blumenthal-Hawley framework](https://www.blumenthal.senate.gov/newsroom/press/release/blumenthal-and-hawley-announce-bipartisan-framework-on-artificial-intelligence-legislation) of September 8, 2023: a licensing regime run by an independent oversight body. It was a framework, not a bill.

The second is inspection before release. The [Artificial Intelligence Risk Evaluation Act](https://www.congress.gov/119/bills/s2938/BILLS-119s2938is.htm) (S. 2938), introduced on September 29, 2025, would have the Department of Energy run an evaluation program.

Under S. 2938, a system trained with more than 10^26 operations could not be deployed until its developer complied, with fines of at least $1,000,000 per day of violation.

The third is liability. The [AI LEAD Act](https://www.congress.gov/119/bills/s2937/BILLS-119s2937is.htm) (S. 2937), introduced the same day, treats AI systems as products and makes a developer "strictly liable" for harm caused by a defective one. Both bills were introduced in September 2025; neither is law.

## Why does AI regulation look like nuclear power regulation?

The comparison is ours, not a lawmaker's, but the shape is the same: license the operator, inspect before operation, and hold the builder liable regardless of care.

That shape works best when the hazard is a large, fixed facility built by a few parties who can be named in advance.

A frontier training run fits that description reasonably well. It needs a very large cluster, which is why every model rule above keys on compute. The EU AI Act [presumes systemic risk](https://artificialintelligenceact.eu/article/51/) above 10^25 floating-point operations of training compute.

What happens after training does not fit it. A model is copied, served behind an API, fine-tuned, and wired into agents by thousands of organizations. The facility-style tools follow the developer, and most of the activity happens somewhere else.

## Which AI rules are actually in force in 2026?

Some model rules are in force. The EU's obligations for general-purpose AI model providers [entered into application on August 2, 2025](https://digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers), and the Commission's enforcement powers, including fines, began on August 2, 2026.

California's [SB 53](https://www.gov.ca.gov/2025/09/29/governor-newsom-signs-sb-53-advancing-californias-world-leading-artificial-intelligence-industry/), signed on September 29, 2025, requires large frontier developers to publish a safety framework and gives the public a channel to report critical safety incidents.

Both are disclosure and risk-management regimes on model developers. The U.S. federal licensing, evaluation and strict-liability proposals are still bills. The table puts each lever next to its status:

<table style="width:100%; border-collapse:collapse; margin:1.5rem 0; font-size:0.95rem;">
  <thead>
    <tr style="background:#f5f5f0; border-bottom:2px solid #2175C5;">
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Layer</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Lever</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Status (November 2026)</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Reaches an agent that swaps models?</th>
    </tr>
  </thead>
  <tbody>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Chips</strong></td>
      <td style="padding:0.75rem;">Commerce Department export licences for H200-class chips to China</td>
      <td style="padding:0.75rem;">In force: <a href="https://bis.gov/press-release/department-commerce-revises-license-review-policy-semiconductors-exported-china">case-by-case review</a> since January 13, 2026</td>
      <td style="padding:0.75rem;">No. It shapes who trains frontier models</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Model (EU)</strong></td>
      <td style="padding:0.75rem;">General-purpose model obligations; systemic risk above 10^25 FLOPs</td>
      <td style="padding:0.75rem;">In force since August 2, 2025; enforcement since August 2, 2026</td>
      <td style="padding:0.75rem;">Only the models in scope</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Model (California)</strong></td>
      <td style="padding:0.75rem;">SB 53 frontier safety frameworks and incident reporting</td>
      <td style="padding:0.75rem;">Signed into law September 29, 2025</td>
      <td style="padding:0.75rem;">Only the developers in scope</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Model (U.S. federal)</strong></td>
      <td style="padding:0.75rem;">DOE pre-deployment evaluation (S. 2938); strict product liability (S. 2937)</td>
      <td style="padding:0.75rem;">Proposed: introduced September 29, 2025</td>
      <td style="padding:0.75rem;">Only the developers in scope</td>
    </tr>
    <tr style="background:#f0f9ff; border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Deployment</strong></td>
      <td style="padding:0.75rem;">Which agents run, what tools each may call, what is screened and logged</td>
      <td style="padding:0.75rem;">Available now to every organization that runs its own agents</td>
      <td style="padding:0.75rem;">Yes. The controls sit around the agent, whatever model it calls</td>
    </tr>
  </tbody>
</table>

## How do chip export controls decide who builds frontier AI?

By deciding who gets the compute. On January 13, 2026, the Commerce Department's Bureau of Industry and Security [said it would review licences](https://bis.gov/press-release/department-commerce-revises-license-review-policy-semiconductors-exported-china) to export Nvidia's H200, AMD's MI325X and similar chips to China case by case.

The conditions are specific: exports must not reduce production capacity available to U.S. customers, the Chinese buyer must screen its customers, and each product needs independent third-party testing in the United States.

That followed a December 8, 2025 presidential announcement allowing H200 sales to approved Chinese customers. That change came from a licence decision at one agency, while the model bills above were still in committee.

The contrast with the model bills is the point. S. 2938 and S. 2937 remain in committee more than a year after their introduction, while a licensing policy issued by one bureau already decides which buyers can assemble frontier-scale compute.

## What does the FBI ARTEX alert show about regulating AI models?

That the harm can sit entirely in the agent. The FBI's [FLASH alert of November 13, 2026](https://www.ic3.gov/CSA/2026/261113.pdf) describes ARTEX, an open-source agentic framework that threat actors have used since at least August 2026.

In the FBI's words, ARTEX "does not include its own" model. It connects to Anthropic Claude, OpenAI GPT and open or Chinese-origin models including DeepSeek, GLM, Grok and Kimi, and the FBI found Chinese models most prevalent in attacker-controlled instances.

Some intrusions went from initial targeting to privileged access in under three hours. The FBI cites exfiltration of personal information on about 67,000 people from South Korean financial institutions, and more than 1,600 exposed ARTEX consoles as of early November.

A licence on one developer does not follow that tool. When one model refuses, the operator switches to another. The original repository was taken down in October 2026, and the FBI says forks, mirrors and a commercial version keep it available.

We wrote about what the alert means for one sector in [the FBI ARTEX advisory for K-12 school districts](/blog/fbi-artex-advisory-k12-school-districts-ai-agents). The regulatory point is broader: the controls that would have mattered sit around the agent, not inside any one model.

## Who is liable when an AI agent causes harm?

Under the AI LEAD Act as introduced, mainly the developer, under strict liability for a defective product. A deployer is treated as a developer if it substantially modifies the system or intentionally misuses it against its intended use.

That is a real shift if it passes. It is also a reminder that the organization running the agent is in the chain. Kenya's draft policy goes further, spreading liability across [the whole deployment chain](/blog/kenya-ai-liability-law-full-chain), from developer to deployer to user.

Either way, the defence is the same: being able to show what the agent was allowed to do, what it was asked, what it answered, and who approved the action. That record exists only if the deploying organization keeps it.

## What can an organization govern while AI laws are still being written?

Its own deployment layer, starting now. The question to answer is not which model is safest this quarter, but which agents exist, what each one may touch, and whether you can reconstruct what each one did.

On ibl.ai each AI agent has its own instructions and its own tool list, switched on or off tool by tool. A moderation prompt screens what a user asks, and a safety prompt screens the model's answer before anyone sees it. Conversations stay reviewable.

Because ibl.ai is model-agnostic, those controls do not change when the model does. A model that becomes restricted, as happened in [the de facto AI regulation episode earlier this year](/blog/de-facto-ai-regulation-government-agencies-2026), is a configuration change, not a rebuild.

Ownership is what makes the record yours. On ibl.ai you own all the code and the data, and you deploy anywhere: your cloud, on-premise or air-gapped, so the evidence a regulator or court asks for sits on infrastructure you control.

ibl.ai is family-owned and operated from New York, NY, and the platform serves 1.6M+ users across 400+ organizations. Agencies can see the deployment options on the [government page](/solutions/government), and the agent layer itself on [Agentic OS](/product/agentic-os).

## Why does owning the AI stack matter?

**ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.**

- **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours.
- **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform.
- **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ.
- **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.

1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.

ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.
