Guide

Runtime security and runtime governance: where one ends and the other starts

Buyers ask for runtime security. Examiners ask for runtime governance. They are two questions about the same call, and the vendor market answers the first far more often than the second.

The numbers show both questions arriving at once. In Andreessen Horowitz's October 2026 survey of 100 CISOs, internal AI agent security led the 12-month buying plans at 45%. In the same survey, 46% named agents that exceed their permissions or get compromised among their top concerns. The first number is a security purchase. The second is a governance problem.

What runtime security answers

Runtime security answers one question: should this call proceed? The control sits in the path between the agent, the model and the tools. It reads the prompt before the model does, reads the response before a person does, and checks each tool call before it runs. A hidden instruction in a retrieved document is caught on the way in. Customer data in an answer is caught on the way out. An agent about to drop a table is stopped before the command executes.

That is the work of an AI security team, and it is necessary. The full definition is on the runtime AI security page, and the wider category is defined under AI agent security.

What runtime governance adds

A regulated enterprise gets three more questions, and none of them is about the attack.

Whose authority was the agent using? An agent works under a person's sign-in. A log that records the person has recorded the wrong actor. See agent authority scope.

Who decided? When a call falls outside what the agent was allowed to do, a named, accountable person has to make the decision. A ticket queue cannot be that person. See human in the loop and the step-up approval guide.

Where is the record? The decision and the evidence of it have to be produced by the same operation, at the time, in a form an examiner can read. See tamper-evident audit log.

Those three are runtime governance. The CISO's guide to runtime governance for AI agents covers them in full.

The two side by side

QuestionRuntime securityRuntime governance
What it decidesWhether this call proceedsWhether the agent had the authority, and who decides when it did not
What it readsThe prompt, the response, the tool callThe same call, plus the identity and the delegated authority behind it
Who asks for itThe security teamThe examiner, the auditor and the board
What it leaves behindA block or an alertA record of the decision and who made it
Where it sitsThe request pathThe request path

One position, two jobs

Both jobs are done from the same place: the request path. A product that sits there can do the first job and stop. Most do. The second job takes an identity for every agent, a way to return a decision to the person who owns the authority, and a record built for someone outside the company to examine.

Run the two as separate products and the enterprise carries two policies and two logs for one call. Run them from one control point and there is one policy, one verdict and one record.

Where the control runs

A security control that inspects prompts and responses sees the most sensitive data the enterprise has. If that inspection happens on a vendor's infrastructure, the control has created a new place for the data to go. Smartflow runs inside the customer's own environment: on-premises, in a private VPC or air-gapped. No request, prompt, response or cached artifact traverses APERION infrastructure.

The test

Ask a runtime security vendor to stop an unsafe call. Most will. Then ask for the name of the person who approved the one that went through, and the record of that approval.

To sort the rest of the market the same way, by the job each category does, see The CISO's Guide to AI Security Categories. It ends with a 90-day plan.

Common questions

Is runtime governance a kind of runtime security? They share one enforcement point in the request path. Runtime security judges the call. Runtime governance adds the authority behind the call and the accountable decision when the call goes past it.

Can a guardrail product provide runtime governance? Guardrails check prompts and responses. Runtime governance also needs an identity for every agent, a route back to the person who owns the authority, and a record of each decision. See AI guardrails.

Which should a regulated enterprise put in place first? Start with the agent that carries the most authority, and put one control in its request path that does both jobs.

More answers are in the AI governance FAQ and the APERION FAQ.

Sources

Figures verified against the source above on November 22, 2026.

The CISO's guide to AI security categories · Runtime AI security · The AI security landscape, mapped by control point · The CISO's guide to runtime governance for AI agents · AI agent authorization · Runtime governance · Runtime AI security, defined · AI agent security · Glossary · All comparisons

Craig Alberino
Craig Alberino
Craig Alberino is the Co-Founder and CEO of APERION, which builds the runtime governance layer for AI agents in regulated enterprises. Inline policy enforcement and identity-bound audit, deployable on premises.

Put this in the path of your own agents.

Policy enforced inline between your agents and every model and tool they reach, with a record bound to the human who owns it.

Request a Demo Read the docs