verifiedagents.ai
All posts

5 min readAI Agents · Cybersecurity

How do you test whether you can trace what your AI agent treats as true?

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

Hero: How do you test whether you can trace what your AI agent treats as true?

TL;DR: Take one thing your AI agents produced last week and trace every claim in it back to a real, working source. Give yourself ten minutes. If you can do it, you can trace what your agents treat as true. If you can't, start a list of every source they read, with a name and a date on each.

Identity checks answer one question. Is this agent allowed to act?

They say nothing about whether the facts behind the action are real. Those are two separate questions, and most of the money so far has gone to the first.

I found the second one the hard way.

Why isn't identity enough?

Because an agent can pass every identity check while acting on something false.

In my lab, I ran a local model on a research task. The output looked right. It had clean formatting and a calm tone, with sources that sounded real. I checked it hours later. Every citation was invented.

By then, other work had been built on top of it. I had to trace back through each decision that used it.

No permission was wrong. Every agent was who it said it was. The information was still fake.

John Kindervag, who created Zero Trust, puts it in three words: identity is crackable. I'd add that even uncracked identity doesn't check the facts. Zero Trust verifies the connection. Whether the content is true is one more check on top.

What does your agent treat as true?

Everything it reads as fact. That includes the knowledge base someone built last year and the folder of shared documents. It also includes the instructions pasted into a settings panel and anything pulled live from the web.

Michelle Savage, my co-author, calls the governed version of that pile the Context Layer. It's the same pile with an owner's name on it and a place where it lives.

In most companies, nobody has ever looked at that pile as one thing. The longer argument is on MassiveScale.AI, in how to run a belief audit.

How does the assessment score this?

Question 18 of the free assessment asks whether you can trace where agent data came from and how it was used. It's part of Data Governance, one of the five ATF elements. The trail from a claim back to its source is called data lineage. Each answer does differently against a ten-minute clock.

Answer

What it says

How the ten-minute trace goes

A

No data lineage tracking.

You can't start. There's no trail.

B

Manual documentation of data sources.

You find a list of sources, but not which one fed this claim.

C

Input source tracking for each agent.

You can name the agent's sources. Matching a claim to one takes digging.

D

Full transformation logging that shows how data was processed.

You follow the claim back step by step, inside the ten minutes.

E

Complete lineage graphs you can see.

You click the claim and the path is drawn for you.

How do you run the ten-minute trace test?

  1. Pick one output from last week. Choose something that reached a customer or a decision.

  2. Mark every factual claim in it. Mark the prices and the dates. Mark the policy statements and the numbers too.

  3. Start a ten-minute clock.

  4. Trace each claim to its source. A source is a link that opens or a record number that exists.

  5. Count what you couldn't trace.

The pass bar is every claim traced inside ten minutes. "According to industry research" isn't a source. A link that opens to a real page is.

Do it on a quiet afternoon. If you can't do it then, you can't do it with an auditor standing behind you.

What do you fix when the test fails?

Build the list the test was missing.

  • Write down every place one agent chain pulls facts from.

  • Put a name and a date on each source. If nobody approved it, it isn't governed.

  • Find the one fact that would hurt most if it were wrong. Check it against your system of record before another agent builds on it.

  • Make every agent show its work. No source means no handoff.

That last rule only holds if a machine can check it. Here's the difference.

Rule as written

Can a second agent check it?

Why

"Never make up a source."

No

An agent that invents a citation doesn't know it did.

"Only use trustworthy information."

No

Trustworthy is an opinion, and the agent grades its own.

"Every fact carries a link that opens to a page on the approved list."

Yes

A second agent can open the link and check the list.

"No working source, no handoff."

Yes

The check happens at the handoff, before the mistake spreads.

I added that last rule to my own lab. A cited page has to open, or the work waits for a person. It's slower. I take that trade every time.

Checking what moves between agents has its own test, in how to verify what one AI agent hands to the next.

Frequently asked questions

How long does the first list take?

About an afternoon for one agent chain. Listing the sources is the slow part.

Who should own the list?

Whoever owns the agent chain. Most of the work is asking who approved each source. Bring security in when you pick the fact to check against your system of record.

How is this different from a knowledge base?

A knowledge base is one source. What your agent treats as true is all of them at once.

How often should I rerun the trace test?

Monthly. Pick a different output each time. For rules about what an agent must never say or do, see testing whether a never-list rule can be checked.

Key takeaways

  • Identity proves who an agent is. It doesn't prove what the agent read was true.

  • Test with one real output and a ten-minute clock.

  • A pass is every claim traced to a source that opens.

  • Write rules a second agent can check.

Question 18 is one of 30 in the free assessment. It takes about ten minutes and scores you on all five ATF elements.

Your agents will keep passing the identity test. What catches up with you is what they believed, and how far it traveled before anyone checked.

See where your agents stand.

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