Congress Just Wrote Four Rules for AI Agents. They Fit Your Business Too

On September 15, 2026, Representatives Mike Lawler of New York and Josh Gottheimer of New Jersey introduced the Stop Rogue AI Act, a bipartisan bill aimed at the AI agents that now run inside government and corporate networks. According to the sponsors' release, the bill would direct the National Institute of Standards and Technology to develop national standards for "discovering, verifying, and controlling AI agents."

Two things need saying before anything else. This is a bill, not a law. It has been introduced, not passed, and nothing in its current text imposes a duty on a private company that does not sell to the federal government. And yet the four capabilities it names are the ones any business should have on the day it lets an agent touch its phone line, its inbox or its customer list. The bill hands you a vendor-neutral checklist a year before NIST finishes writing the standard. This article is information, not legal advice.

What the bill actually asks for

The release lists four things the standards would teach organizations to do. Find and track every AI agent operating on their systems, in a continuous, easily readable inventory. Verify who built and operates each agent, using verifiable identity and provenance, "not just a vendor's word." Monitor agents in real time, including detecting prompt injection, data theft, and agents behaving outside their approved limits. And allow, deny, or revoke an agent's access, actions and interactions at any time, "so people have the final say."

The bill text goes further than the summary. NIST, in coordination with the NTIA, would have to publish the standards within one year of enactment and update them annually. The list of required capabilities includes runtime monitoring that catches prompt injection, data exfiltration, anomalous tool invocation and "behavioral drift from an AI agent's approved operational baseline," plus tamper-evident logs of material agent actions.

Minimum organizational requirements are spelled out too. The inventory must be machine-readable and use standardized, vendor-agnostic naming. Agent identity must be cryptographically verifiable at both the network and the application layer. And an organization may not rely "solely on self-attested or single-provider assertions" to establish who an agent is.

Who it binds, and who it does not

The enforcement path runs through federal procurement. Within 18 months of the standards being published, the Federal Acquisition Regulatory Council would propose revisions so that contractors and agencies buying or deploying AI agents comply with the standards. Contracts signed before enactment are exempt. Within 180 days of that proposal, the Office of Management and Budget, working with CISA, would issue guidance to agencies. If you do not hold a federal contract, none of this reaches you directly.

The definition of an AI agent is worth reading closely, because it tells you whether the tool you run would even count. The bill describes a software system that uses an AI model to perceive, plan or make decisions, autonomously interacts with other systems or people on behalf of a person or organization, and involves "minimal or no human interaction beyond its initial direction." A phone assistant that books appointments on its own fits that description. An agent that drafts and sends follow-up texts fits it. A chatbot that only answers questions and never takes an action sits closer to the edge.

Rule one: know every agent you run

The cheapest control in the bill is a list. Which agents run in your business, who owns each one, what data and tools it can reach, and when someone last reviewed it. For a five-person company this is a spreadsheet, and "continuous" means it gets updated the same day a new tool is switched on, not at the end of the quarter. Most small businesses that run two or three AI tools cannot answer the question "what can that thing actually read and send" without opening a settings page. The list fixes that.

Rule two: verify who built it, not just who sold it

The bill's phrase is "not just a vendor's word," and it lands hardest when you buy through a reseller or an integrator. The questions are plain. Who built the model. Who operates the runtime where your data flows. Can they show that, or only say it. The bill wants identity that can be checked cryptographically and forbids leaning on a single provider's self-description. You do not need the cryptography to ask the questions, and a vendor who cannot answer them is telling you something.

Rule three: watch it while it works

Prompt injection sounds technical until you see the plain version. An inbound email tells your agent to forward the customer list to an outside address, and the agent, which was built to follow instructions, follows that one. Data exfiltration is the agent quietly exporting records it was never asked to touch. Behavioral drift is the agent that used to book appointments starting to send discounts nobody approved. The bill asks for monitoring that catches these in real time. For a small operation that means an agent with a log you actually read, and alerts when it does something outside the pattern.

Rule four: keep the off switch and the record

Two questions to ask before signing any agent contract: can you cut its access in one minute, and can you show what it did yesterday. The bill requires the ability to allow, deny or revoke agent access at any time, and tamper-evident logs that stay portable and accessible to the organization that deployed the agent. An agent you cannot switch off is not a tool, it is a liability with a subscription fee.

NIST is already on this

The bill does not start from zero. NIST launched its AI Agent Standards Initiative in February 2026, built around three pillars: industry-led standards, community-led protocols, and research into agent authentication and identity. The same page lists a request for information on AI agent security and an NCCoE concept paper on agent identity and authorization. What the Stop Rogue AI Act would add is a dated deliverable and a procurement hook.

A Monday checklist

Write down every agent you run. Next to each one, write who built it, who operates it, what it can read and what it can send. Confirm you can revoke its access yourself, without a support ticket. Confirm you can read its log. Put a review date on the sheet. That is the whole bill, translated for a business with no compliance department.

PATech builds and runs AI agents for small businesses, including voice agents that answer the phone line, and designs them so the owner keeps exactly these four controls: an inventory, a verified operator, a monitored runtime and a kill switch with a log. Talk to PATech if you want to see what that looks like on your own phone line.

The bill may pass or it may not. The four rules cost nothing today.

This material is for informational purposes only and is not legal, financial, or other professional advice.