Blog Wiz John Dwyer 1

Part 1 of 4

You Can't Secure the AI You Don't Know You Have

In my recent series, AI Is Moving Faster Than the Operating Model Around It, I spent a lot of time on authority.

How much authority does an AI system have? What data can it access? Can it recommend an action, make a decision, or actually execute one? What happens downstream when it is wrong? I still think those are some of the most important questions organizations should be asking about AI, but there is a problem that comes before all of them.

Do you actually know where the AI is?

I think a lot of AI security conversations start too far downstream. We immediately start talking about prompt injection, model security, guardrails, AI governance, or how to safely deploy agents. Those are important conversations, but you cannot govern authority you cannot see.

And I suspect many organizations have a much less complete understanding of their AI environment than they think they do.

An AI inventory is not a list of applications

When someone says "AI inventory," it is easy to imagine a spreadsheet containing approved applications:

Microsoft Copilot.
ChatGPT Enterprise.
Gemini.
AWS Bedrock.

That is useful, but it is not an AI inventory. Modern AI applications are systems. There may be models, agents, datasets, pipelines, tools, identities, APIs, MCP servers, guardrails, application endpoints, and the underlying cloud infrastructure involved in delivering a single business capability. Some of those components may have been deliberately deployed by a central AI team while others may have been created by developers (and some may exist inside SaaS platforms). In my experience, AI is often running inside a workload that nobody thinks of as an "AI application" at all. This is one of the reasons I think Wiz is becoming such an important technology for AI security.

Wiz does not limit AI discovery to a list of sanctioned applications. Its AI security capabilities can identify agents, models, guardrails, MCP servers, tools, pipelines, datasets, deployments, gateways, application endpoints and even agentic identities.

The discovery mechanisms are equally important. Wiz can build visibility from cloud and AI connectors, workloads, data, source-code repositories, CI/CD pipelines, logs, attack-surface scanning and runtime sensors.

Wiz refers to the resulting inventory as an AI bill of materials, or AI-BOM.

I like that concept because it forces us to stop thinking about AI as a product somebody purchased and start thinking about the collection of technology that actually creates the AI capability.

Discovery is necessary. It is not the outcome.

Discovery is important, and inventories are useful, but this is also where I think we need to be careful. Security teams have spent decades building bigger inventories and producing more findings. The goal cannot simply be to create an enormous inventory of AI resources and congratulate ourselves for finding them. Finding an AI agent is useful, but understanding what that agent can cause to happen is much more useful.

What security teams really need to know about an AI system is

Who can invoke it?

Which identity does it operate as?

Which tools can it use?

What data can it access?

Can it change the environment?

Is that data sensitive or regulated?

What guardrails exist?

What happens to the business if the agent does something incorrectly?

Those questions sound much closer to traditional security engineering than some of the AI-security discussion happening today. The good news is that we do not need to throw away everything we already know about security simply because the technology is new.

Identity, permissions, data, and architecture remain important, as do exposure and vulnerability management. The difference is that AI introduces new relationships among those things and, increasingly, gives software the ability to make decisions and take actions across them.

This is another area where Wiz gets interesting. Its Security Graph can correlate exposure, identity, data and vulnerability information to identify combinations of risk and attack paths involving AI systems rather than treating each finding as an isolated problem.

That context matters much more to me than another vulnerability count.

Visibility is not the same thing as risk

Imagine that we discover two AI agents: 

One summarizes internal documentation using a read-only identity and the other can access customer data, invoke external tools and modify production resources.

Those should not create two equal entries in an AI-risk register simply because they are both agents. The underlying technology may be very similar. The risk is not.

This gets back to a principle from the previous series: security should care less about how sophisticated the AI appears and more about what the system can cause to happen, but technology can only take that analysis so far. A security platform can tell us that an agent has access to a particular dataset. It cannot independently determine how important that dataset is to the business. It can identify an exposed AI service, but it cannot decide the organization's tolerance for the business process supported by that service. It can identify a control failure, but it cannot determine by itself how that failure intersects with the organization's regulatory obligations, risk appetite, architecture, compensating controls, and mission.

That requires expertise.

This is where I think Coalfire and Wiz become much more interesting together

At Coalfire, we sit in an unusual position because we work across the security lifecycle. We have people who design and build environments.

People who assess them.

People who try to break them.

People who defend them.

And people who understand the regulatory and compliance obligations surrounding them.

All of those disciplines look at the same technical environment differently. An engineer may look at an AI agent and think about architecture, identity and trust boundaries. An offensive-security practitioner may ask how those relationships could be abused. A defensive-security practitioner may ask what telemetry exists to identify that abuse and how we would respond. A compliance practitioner may ask what regulated information is involved, which control requirements apply, and what evidence is necessary to demonstrate that the risk is being managed.

None of those perspectives is sufficient by itself. Together, they start turning technical visibility into an understanding of risk. That is the part of AI security I think organizations are going to struggle with over the next several years.

Not because they lack security products. Because they will have an enormous amount of new information about AI systems and will need to determine what deserves attention now.

The platform and the expertise solve different problems

This distinction is important. I do not think the answer is to perform a point-in-time AI assessment, create an inventory and come back next year. AI environments are changing too quickly for that. The inventory begins to become stale the moment the assessment ends. We need continuous technical visibility, but continuous visibility by itself can create a different problem: another platform producing information that an already constrained security organization now has to interpret, prioritize, remediate and operate.

That is why I don't view technology and services as substitutes for one another here. Wiz can provide something extremely valuable: continuous ground truth about the AI environment and the relationships surrounding it.

Coalfire can help organizations determine what that ground truth means in the context of their architecture, security program, compliance obligations and business risk, and then help them do something about it.

The combination creates a much more useful cycle:

Discover what exists.
Understand what it can do.
Determine what actually creates risk.
Fix what matters.
Then continue watching as the environment changes.

That is very different from deploying another dashboard.

Start with visibility, then earn the right to do more

I also think this gives organizations a practical way to approach AI security without trying to solve everything on day one. Start by understanding the environment. Find the AI services, models, agents, identities, data, tools and infrastructure that already exist. Understand the relationships. Identify obvious exposure and configuration problems. Identify some quick wins. Then mature from there.

Wiz can surface missing or weak guardrails, excessive agency, sensitive-data relationships and risky access paths. As the program matures, capabilities like Wiz Code and Wiz Defend can extend that visibility further into development and runtime behavior.

That progression makes sense to me.

We do not need to pretend an organization can establish a mature AI-security program overnight.

But we should know where we are starting.

You cannot secure what you cannot see

The first question I would ask a CISO about AI security is not:

"Do you have an AI policy?"

It is:

"Can you show me where AI exists in your environment?"

Then I would keep going.

What can those systems access?

What can they do?

Which identities do they use?

What business processes depend on them?

Which systems or data could they affect?

Who owns the risk?

If we cannot answer those questions, our AI governance may look good on paper while a completely different AI environment is developing underneath it.

AI security starts with visibility.

But visibility only becomes valuable when somebody can turn it into context, decisions and action.

That is where I think the real opportunity exists.

And it is where the combination of Wiz and Coalfire starts to become much more powerful than either technology or consulting delivered in isolation.