What ‘AI taxonomy’ looks like at Entrust

When Adam Dimopoulos, CISO at digital security company Entrust, first assumed his role, one of his first orders of business was establishing a formal classification system for the company’s various AI systems.

“I said, ‘Look, we’re not going to be able to implement these guardrails that we’ve predefined unless we classify the different types of AI systems, because it’s going to be a different set of guardrails,’” Dimopoulos told IT Brew.

Dimopoulos said the categorization effort, which he termed “AI taxonomy,” was important because AI tools vary in authority models from vendor to vendor, requiring different governance structures. In May, Gartner found almost half (40%) of enterprises will demote or decommission autonomous agents due to governance gaps caused by uniform governance policies.

“This is really foundational or important because you can’t apply governance as a single blanket uniformly across all the different types of AI systems,” Dimopoulos said. “You have to define what the different AI systems are, how the AI systems access your data, what they’re going to do, and who the personas are.”

ArmorCode Director of AI Matt Sayar told IT Brew that uniform governance policies for humans and AI tend to fail for the same reason: different departments, or AI tools in this case, have specific needs: “You don’t necessarily need your HR department to access your live servers in production for financial data. Same thing with AI.”

Break it down. These days, Entrust classifies AI into three different categories: AI systems that consume and produce information for the general workforce (e.g., enterprise LLMs); AI that influences production systems (think coding assistants); and agentic systems.

Dimopoulos uses this breakdown to determine appropriate guardrails for specific AI tools. When it comes to governing enterprise LLMs and other similar AI systems, for example, the focus is largely on data protection.

“We have a rule that says you cannot use a personal login to a public LLM to input corporate information,” Dimopoulos said. “And then we actually have capabilities to monitor if people do that.”

For coding assistants, protections are focused on the software development life cycle (SDLC) and sandbox testing. “If they’re just working from a very small approved list [of AI tools], they can’t innovate,” Dimopoulos said. “So, we need those isolated developer sandboxes where they can experiment with new AI with less of a concern for some of the risks and then we can build the protections into the pipeline over time.”

Governance of agentic systems looks similar to the principles of identity governance, said Dimopoulos. For example, an AI agent at Entrust can never have more authority than the “requesting principal”: “That means if my identity is the owner of an agent and I only have the ability to read financial data, that agent should not have the ability to approve that financial data,” he said.

Taxonomy pays off. Dimopoulos has seen the benefits of his work. Prior to establishing AI taxonomy, Dimopoulos said, the company’s full enterprise approval process often took too long for developers. Now, developers can quickly experiment with AI coding assistants in sandbox environments.

“If you need to move it out of the sandbox into a more pre-staging environment to test with some live data or production data, then you need to build in these specific SDLC pipeline controls,” Dimopoulos said.

“Those controls would not apply to an LLM, but they apply very specifically to that coding assistant. So, that’s one way that we saw immediate value out of introducing this model,” he added.

For companies with blanket AI governance policies, Dimopoulos suggested creating standards to classify different systems: “Standards are good ways to check yourself and then try to enforce some actual control overlays.”