Guide

The CISO's Guide to AI Security Categories: Stop Sorting, Start Governing

The CISO's Guide to AI Security Categories: Stop Sorting, Start Governing

Every analyst, investor and conference is publishing a map of AI security, and the agents those maps describe are already running in production. This guide sorts the categories into the two jobs behind them, shows where 2027 budgets are going, and lays out how to act inside 90 days.

Everyone is drawing the map

In late October, Andreessen Horowitz published results from a survey of 100 CISOs at companies with more than 1,000 employees, with 2027 buying plans broken out by AI security category. a16z, citing Jake Epstein's Black Hat USA 2026 vendor landscape, notes that some categories include as many as 43 vendors. Gartner published its first Magic Quadrant for AI Governance Platforms in June 2026. We published our own map in September, organized by the control point where each product intervenes.

Add the market maps, the analyst briefings and the conference floor, and a CISO is looking at hundreds of vendors describing themselves in nearly identical words: autonomous, agentic, runtime, control plane. Sorting them has become a job of its own.

Two jobs behind one market

a16z found two jobs driving the 2027 security agenda. Most category confusion comes from mixing them.

Job one: secure the AI the company already runs. Employees use generative AI tools, developers use coding agents, and agents reach documents, source code and production systems through MCP servers and integrations. The question here is what those agents may do, who decides when they go further, and what proves it afterward.

Job two: use AI to do the security work. Agents that triage alerts, run attack tests, find vulnerabilities and draft fixes. The question here is how much of the security team's work a machine can take on, and how much authority it gets to act.

The table is our sorting of the categories buyers now see, by job.

Category What it does Job
Internal AI agent security Controls what agents may do as they act: tool calls, data access, model calls One
AI usage and governance Policy, inventory and oversight for the AI tools employees use One
AI endpoint prevention Controls software, developer tools and agents on devices One
AI data security Classification, DLP and access governance for data AI can reach One
AI identity governance Access reviews, provisioning and least privilege for people, workloads and agents One
AI GRC and vendor risk Questionnaires, evidence collection and vendor assessments Both
AI SOC Triage, investigation and response Two
AI detection engineering and threat hunting Writing and tuning detections, hunting for intruders Two
AI offensive security Autonomous testing, red teaming and attack simulation Two
AI exposure and vulnerability remediation Prioritizing, fixing and verifying fixes Two
AI code security Code review, vulnerability detection and automated fixes Two
AI email and social engineering defense Phishing triage and deepfake and impersonation defense Two

Every job-two category is a new fleet of agents with authority inside your environment. An AI SOC agent that isolates a host, a remediation agent that ships a fix, a testing agent that runs against production: each is a job-one problem the day it goes live. Job two creates job one.

The first row of that table goes by several names. The a16z survey calls it internal AI agent security. Vendors call it runtime security. A regulated firm needs the version that adds a named decision-maker and a record, which is runtime governance. The definitions are under runtime AI security and AI agent security. Runtime security and runtime governance draws the line between the two.

Where the money is going

From the a16z survey of 2027 plans, published October 28, 2026:

  • 87% expect security spending to go up, mainly because attackers now use AI and because their own AI rollouts widen what they have to defend.
  • Every one of the seven initiatives chosen most often for 2027 involves AI. When the same CISOs looked back at 2026, exposure and vulnerability management and SIEM led.
  • Internal AI agent security tops the 12-month buying plans a16z reports: 45% intend to buy or upgrade within a year. Offensive security is next at 42%, then AI usage and governance at 38%.
  • Production use still trails intent. AI SOC tools are live at just 18% of these companies.

The buying window is open now. In the top two categories, more than four in ten buyers plan to act within the year.

Why buyers are stuck

The paralysis has good reasons behind it.

The claims sound the same. Dozens of vendors promise more autonomous security, quicker investigations and less work by hand, and a16z notes how alike their descriptions sound.

Evaluation does not scale. A CISO at a global apparel company, quoted by a16z, faced seven or eight candidates and took only three into hands-on trials. Running a proof of concept on all of them, he said, was "not humanly possible."

The products move mid-evaluation. A vendor ships a release, a new entrant appears, a model upgrade resets the baseline, all before the proof of concept ends.

Point products pile up. One security leader quoted by a16z does not want to keep adding point products, and 44% of respondents turn first to features in a platform they already run.

Each of those is a sound reason to be careful. Together they produce a decision that keeps slipping to next quarter.

The clock does not wait for the evaluation

While the shortlist is debated, the environment keeps changing.

  • Attacks are faster. Bain puts a typical attack sequence run with AI at about eighteen hours, down from roughly four weeks.
  • The agents are already deployed. In the a16z survey, 92% have employee generative AI tools in production, if only in a limited way. Coding assistants or agents are live at 83%. At 55%, MCP servers or similar links between agents and applications are already in production.
  • Access grows without a decision. Software cleared as an ordinary app can arrive with agent features in a routine update. Employees install extensions and integrations as they go. Each new assignment widens what an agent can reach.
  • The worries are already concrete. Among their top concerns, 47% of CISOs list staff sending sensitive data into AI tools, and 46% list agents that overstep their permissions or get compromised.

Every month spent sorting vendors is a month those agents can act with nothing between their decision and its consequence.

How to act inside 90 days

Start with the agents that can do the most damage, on a date you set. The rest of the stack can follow.

  1. Split your list by job. Decide which job is your first priority. In the a16z survey, SOC automation was ranked first most often. Other teams led with offensive security, remediation or protecting their own AI. Either answer works. No answer does not.
  2. Inventory what is already running. Agents, MCP servers, coding assistants, extensions and the integrations behind them. Give it two weeks. An imperfect inventory you act on beats a complete one you finish in March.
  3. Find the agent with the most authority. The one that can move money, change production or reach customer records. Start there.
  4. Match the control point to where that agent runs. An agent on a laptop, an agent inside a SaaS platform and an agent running on your own servers need different controls. Our control-point map shows what each position can and cannot see.
  5. Write down its authority, and who decides past it. For each high-consequence agent: what it may do on its own, what needs a person, and which person. If nobody can answer, that is the finding.
  6. Run one scoped proof of concept, with three vendors at most. Use your own traffic. Set the exit criteria before it starts: what the product stopped, what it sent to a person, and what record your examiner would receive.
  7. Buy for composition. Assume you will run more than one control point. Require every tool to write its records into your SIEM, so the evidence lives in one place you control.
  8. Put the date on the calendar. Ninety days from today, the highest-authority agent in your environment runs under a control you chose.

Questions that cut through any category label

  • Where does this product act, and what can it not see from there? Every vendor answers the first half. The evaluation happens in the second.
  • When an agent goes past its authority, what happens? A block, an alert, or a decision by a named person?
  • Who approves, and how is that person verified at the moment they approve?
  • What does the examiner receive for one action? Ask for a sample record.
  • Can it run where our data has to stay? Ask about on-premises and air-gapped operation by name.

Where APERION sits

APERION works in job one: internal AI agent security for regulated enterprises. Smartflow sits inline on the model request path and the MCP tool path, on the customer's own infrastructure, including air-gapped. Policy is enforced before each call routed through it executes. When an action needs a person, it pauses until someone with approval rights decides, and step-up can require that person to re-verify their identity first. Every decision is recorded and written to the customer's SIEM.

Our position is owned intelligence, governed from the first call. Identity proves the person. Access scopes the authority. Runtime enforces that authority on each call and pauses any action that goes past it until a person decides. Audit proves who decided. That chain is the product.

We do not run a SOC, test applications or scan code. Several products in the categories above compose with ours, and a serious program will run more than one.

Start this week

Pick the agent in your environment with the most authority. Find out what governs its next call, and who decides when it reaches past what it was given. If the answer is nothing and nobody, you have your first project, and your date.

Common questions

What are the main AI security categories? They divide into two jobs. Securing the AI a company runs covers internal AI agent security, AI usage and governance, endpoint controls for AI tools, AI data security and AI identity governance. Using AI to do security work covers AI SOC, detection engineering, offensive security, vulnerability remediation, code security and email defense. AI GRC serves both.

What is the difference between AI for security and security for AI? AI for security uses agents to do security work, such as triaging alerts or testing systems. Security for AI governs what a company's own AI and agents may do. The first creates a need for the second, because every security agent is an agent with authority.

Which category should a CISO address first? Start with the agent that carries the most authority in your environment, then choose the category that governs it. Internal AI agent security topped the 12-month buying plans a16z reported.

How many vendors belong in an AI security proof of concept? Three at most, tested on your own traffic, with exit criteria set before the test starts.

Platform or startup? Many CISOs check their existing platform first. In the a16z survey, 47% said they would go to a startup for something their current vendor cannot do. Map the gap first, then decide.

Is runtime security the same as internal AI agent security? They describe the same control: enforcement on each prompt, response and tool call while an agent works. Runtime governance adds whose authority the agent used, who decided, and the record. See runtime AI security.

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

Sources

Figures verified against the sources above on November 6, 2026.

The AI security landscape, mapped by control point · The CISO's guide to runtime governance for AI agents · What is an AI control plane? · AI agent authorization · Step-up approval for AI agents · The AI inventory guide · Runtime governance · All comparisons · Runtime security and runtime governance · Runtime AI security · AI agent security · MCP security · The enterprise AI gateway guide · AI governance FAQ · Glossary

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