verifiedagents.ai
All posts

4 min readAI Agents · Cybersecurity

How do you measure an AI agent's blast radius before the inventory is finished?

By Josh Woodruff, Founder & CEO, MassiveScale.AI | Founding Chair, Agentic Trust Framework at the CSAI Foundation

Hero: How do you measure an AI agent's blast radius before the inventory is finished?

TL;DR: You can measure blast radius without a finished inventory. Name your five highest-impact AI agents and score each one on five controls: its own identity, a default ceiling, a fast stop, a log it can't edit, and one route out. Each missing control widens the radius. Then report coverage across the agents you know about.

Security has a saying that's held for twenty years. You can't protect what you don't know is out there. I've said it on stage.

Then in August 2026, the same shift came up on three of my security calls inside eight days. Contain first. Count second. Attackers using AI can find a weakness and use it faster than a discovery project finishes.

So the measuring has to start before the counting ends.

What is an AI agent's blast radius?

It's how much damage one agent can do before somebody stops it.

Four things set the size. How far the agent reaches. How fast it acts. How bad the worst action is. How hard that action is to undo.

None of those need a full inventory. They're facts about one agent and the place it runs. The case for this order is on MassiveScale.AI, in contain first, count second.

How does the assessment score this?

Question 24 of the free assessment asks how you limit damage if an agent is compromised. It belongs to Segmentation, one of the five ATF elements. Each answer leaves a different kind of proof.

Answer

What it says

What you can measure

A

No blast radius consideration.

Nothing. Nobody has asked the question.

B

General least privilege principles are followed.

A policy, with no number per agent.

C

Documented blast radius analysis per agent.

A score for each agent, on paper.

D

Technical controls actively limit each agent's blast radius.

A score for each agent, plus tests that show the controls hold.

E

Full containment, with cascade prevention between agents.

A score that also covers what one agent can do to another.

Which agents do you measure first?

Ask your engineers two questions. Which agents touch money? Which agents can change something a customer sees?

You'll get about five names. Those five are the program. You don't need a discovery scan to find them.

The pull runs the other way, toward easy agents that close fast and look good on a slide. That gives you a clean report and no protection.

How do you score one agent?

Check five controls. Score one point for each one that's missing. Zero is the best score.

  1. Its own identity. It doesn't share a service account or run as the person who launched it.

  2. A default ceiling. By default it gets no production data and no write access to a system of record.

  3. A fast stop. Someone can shut it off without a change ticket, and they've tested that.

  4. A log it can't edit. The log shows what the agent did and under whose authority.

  5. One route out. It reaches outside systems through a single path you control.

A score of zero means a mistake stays small. A score of four or five means one bad action can travel as far as the agent's access goes.

The first control is the one I check hardest. An agent that runs as a person has that person's reach. Afterward, you can't tell their actions apart.

What do you report while the inventory is still running?

Report coverage. Skip the total.

A coverage number sounds like this: of the agents we know about, this share run under their own identity. It's honest, and you can improve it. Leaders will accept a partial number with a trend line.

A total sounds like "we have 412 agents." You'll be defending that number the day it moves.

The inventory keeps running in the background. It just stopped being the gate.

Frequently asked questions

Isn't "you can't protect what you don't know" still true?

It's true. It's just a poor place to start now. Discovery runs slower than agent creation, so waiting on it means the controls never ship.

How is this different from least privilege?

Least privilege limits what one agent may do. Blast radius measures what happens when that limit fails. To check the limit itself, use the ten-minute test of what an AI agent can reach.

Does a small blast radius help against prompt injection?

It limits the damage. It doesn't prevent the attack. What counts is what the agent can reach after it's been fooled.

How do I prove the fast stop works?

Test it on purpose. I walk through that in testing whether you can stop an AI agent before it acts.

Key takeaways

  • Blast radius is the damage one agent can do before somebody stops it.

  • Start with the five agents that touch money or change what customers see.

  • Score five controls per agent. Each missing one adds a point.

  • Report coverage with a trend. Leave totals out.

Question 24 is one of 30 in the free assessment. It takes about ten minutes and scores you on all five ATF elements.

I spent years telling people to find everything first. The attackers got faster, so the order had to change.

See where your agents stand.

The free assessment takes ten minutes and scores you on the five elements of the Agentic Trust Framework.