AI AGENT SECURITY / BUYER GUIDE

Choose controls for the actions that matter.

An AI agent security platform helps teams discover, monitor or govern agent activity. Choose one by the actions it can observe and control, where it connects, and the evidence it produces. Start with your workflow, then evaluate the controls.

AI agent security approaches: what to evaluate before choosing a platform
ControlQuestion it answersWhere to check coverageEvidence to request
Model and harness controlsIs this content or workflow permitted?Prompt, output and tool boundaries supported by the implementation.Test indirect inputs and check whether execution can bypass the boundary.
Identity and access controlsWhich identity may access this resource?Human delegation, agent credentials and target permissions.Scoped identities, credential lifetime and target authorization records.
AI, MCP and API gatewaysIs this routed request permitted?Traffic actually sent through the configured gateway.Route inventory and tests of direct or alternate connections.
Agent sandboxesWhat can this workload reach or change?Process, filesystem and network boundaries configured for the workload.Isolation policy, egress rules and escape or bypass test scope.
Independent execution controlShould this proposed action run in its current context?Supported control point before the target changes.Policy, context, verdict, approval scope and correlated target outcome.
Observability and security monitoringWhat activity did the instrumentation record?Instrumented calls, exported events and connected system logs.Trace continuity, collection delay, missing paths and investigation records.

START WITH ONE WORKFLOW

What does your agent actually do?

List the agent, initiating user, intended task, tools, target systems and credentials. Separate reads from writes, exports, payments, shell commands and delegation. A product that inspects a model request may not see a direct database operation.

Identify which paths need observation and which need a decision before execution. Include permitted work in the test set. A control that blocks everything is not a useful success criterion.

Use the 12-point AI agent security evaluation checklist to record scope, expected results and evidence. It is a worksheet, not a completed benchmark or certification.

ARCHITECTURE

Where should the decision be enforced?

The enforcing component should be outside the agent’s ability to alter the authorization decision, with a supported interception point close enough to check the actual proposed operation before it executes. Physical proximity alone is not a security guarantee. Check control ownership, credential paths and alternate routes.

A map of where 4 security approaches sit relative to one agent. 1: Harness and model controls sit inside the agent and are best at content safety, PII in text and jailbreak resistance. 2: An agent sandbox, for example NVIDIA OpenShell, draws a boundary around the agent's process and is best at limiting what the agent can reach. 3: Ackuity sits beside the agent, in its execution path, in the Goldilocks zone: not inside the agent, not at the gateway. It is best at deciding whether each action, right now, should run. 4: AI gateways sit at the perimeter on the LLM, MCP and API calls routed through them and are best at auth, rate limits, central policy and logging. Other actions cross the perimeter without a gateway. Most teams run more than one layer.2AGENT SANDBOXAgentMODEL + HARNESS1Harness andmodel controlsGOLDILOCKS ZONE3Ackuitybeside the agent,in its execution pathnot inside the agent,not at the gatewayPERIMETER4AI gatewayLLM·MCP·APIOTHER ACTIONSTARGETS1INSIDE THE AGENTHarness andmodel controlsBEST ATcontent safety, PIIin text, jailbreaks2AROUND THE PROCESSAgent sandboxfor exampleNVIDIA OpenShellBEST ATlimiting what anagent can reach3BESIDE THE AGENTAckuityexecution control,Goldilocks zoneBEST ATwhether this action,right now, should run4AT THE PERIMETERAI gatewayson calls routedthrough themBEST ATauth, rate limits,policy and logging
Four places to put AI agent security: harness and model controls inside the agent, a sandbox around its process, Ackuity beside it in the execution path, and AI gateways at the perimeter. Each layer is best at a different job, and most teams run more than one.

Model controls, identity, gateways, sandboxes and execution controls can complement each other. Compare their configured coverage rather than assuming that a category name proves a capability.

DEPLOYMENT FIT

Can it connect to the agents you run?

Ask for the supported route, prerequisites, configuration owner, data flow and demonstrated action coverage for your specific platform and version.

  • Kubernetes agents: inspect how the sidecar or other enforcement component intercepts calls and handles direct connections.
  • Hosted or low-code agents: identify the API or connector control point. A sidecar-only answer is insufficient for a platform where you cannot install one.
  • Existing traces: confirm which events can be collected and when. Event ingestion establishes visibility, not inline enforcement.

Ackuity, the Agent Execution Control Switch, supports a sidecar, API integration for Microsoft Copilot Studio, and observe-only event pull. The connection paths have different coverage. The OpenShell integration is in development. Request a scoped evaluation of the route you will use.

ACTION CONTEXT

Can it distinguish the same call in different situations?

A delete operation could be an authorized cleanup, an unfamiliar bulk change requiring review, or an action outside a hard policy boundary. The operation name alone cannot settle that question.

Ask how the product connects the user, agent, authorized task, target, tool and earlier activity. Separate trusted system records from agent-supplied assertions. Do not assume access to hidden model reasoning.

Ackuity assembles the Agent Security Context Graph across six dimensions and 29 signals. Vinod's SQL example explains why context can change a verdict. The scenario illustrates the decision model; it is not a customer test result.

ACTION COVERAGE

Which routes will your evaluation miss?

Trace the complete path from proposed action to target. Test shell, SQL, HTTP, MCP, memory and delegated work where those paths exist in your environment. Record excluded or unobservable routes explicitly.

In Ackuity's 11-action routing model, 3 action types pass a gateway and 8 do not. Real routing depends on the deployment. A database operation exposed as an MCP tool may cross an MCP gateway; a direct database connection may not. Use a route inventory rather than the action label alone.

Review the incident and control library for sourced examples and testing questions. An incident illustrates a risk, not proof that a specific product would have stopped it.

AUTHORIZATION & APPROVALS

Does an approval authorize this exact action?

Require the execution component to check an approval against the actor, operation, target and parameters. Establish expiry, replay protection and what happens if the request changes. A conversational confirmation is not enough to prove authorization at execution.

Keep hard policy limits separate from discretionary review. An approval must not override an unconditional prohibition. Within permitted scope, an unfamiliar or high-impact operation may need human judgment.

See the scoped human approval runbook and OWASP's transaction authorization guidance.

FAILURE & OPERATIONS

What happens when the control cannot decide?

Test timeout, disconnection, retries, concurrency and actions already in progress. Record fail-open or fail-closed behavior per path. Verify how temporary access expires, credentials are revoked and operators contain an incident.

A kill switch remains useful for emergency containment. Ackuity's control switch provides per-action responses: Allow, Constrain, Human in the loop, Block and Terminate. Evaluate the available scope and recovery behavior instead of treating every implementation called a kill switch as identical.

Measure decision time separately from end-to-end workflow time, and count mistaken blocks on authorized work. Agree on data retention, access, operational ownership and commercial terms before the evaluation. Compare kill switch and control switch.

EVIDENCE BEFORE PURCHASE

What proves an action was allowed or blocked?

Use an isolated target, synthetic records and an approved test identity. Keep these three records together:

  1. Proposed operation: actor, task, target, parameters and correlation ID.
  2. Control decision: policy version, context, verdict and any approval scope.
  3. Target outcome: independent audit record, resource state or transaction result.

Include a permitted action, a prohibited action and a reviewed action. Re-test alternate paths and changed parameters. A blocked decision record alone does not prove that no other path reached the target. Missing target logs mean the outcome is unknown.

Distinguish completed tests from product descriptions, demonstration screens and roadmap features. Keep a record of the version and configuration tested. The downloadable worksheet helps your team compare evidence without inventing a single vendor score.

OPERATING MODEL

Can you observe without claiming prevention?

Start by mapping activity if you do not yet know the agents and routes in scope. AI agent security monitoring connects tool calls and downstream records so investigators can understand what happened. An enforcement evaluation is a separate step.

The 30-day Agent Exposure Assessment reviews agreed telemetry and delivers findings and priorities. It is not an inline-control proof of concept. Agree separately on any controlled enforcement tests.

WHERE ACKUITY FITS

What should you evaluate in Ackuity?

Ackuity is the Agent Execution Control Switch. It checks proposed actions using independent context, rules, behavioral baselines and small language models. No LLM generates the authorization verdict.

The Goldilocks zone means not inside the agent, not at the gateway: beside the agent, in a supported execution path. Evaluate the actual interception point, credentials and bypass behavior for your deployment.

Request evidence for the workflow you intend to protect. The sidecar provides the deepest coverage; hosted-platform APIs and event pull have distinct scope. No single successful test establishes universal action coverage.

METHOD & SOURCES

What is this guide based on?

This is Ackuity's buyer evaluation method, not an independent vendor ranking. It combines our approved architecture descriptions with public security guidance. The criteria are questions to test, not claims that every product in a category passes or fails.

Sources checked October 24, 2026. Source references do not imply endorsement. NVIDIA and OpenShell are trademarks of NVIDIA Corporation.

QUESTIONS PEOPLE ASK

Frequently asked questions.

How do I choose an AI agent security platform?

Start with the agents, actions and routes you need to protect. Compare deployment fit, action coverage, context, authorization, failure behavior and independent evidence of the target outcome. Use the same scoped tests on each candidate.

Is agent security monitoring the same as blocking an action?

No. Monitoring records and analyzes observed activity. Blocking requires a supported enforcement point before the target operation executes. An event-pull integration does not establish inline prevention.

Does an AI agent security platform replace identity controls or sandboxes?

No single category should be assumed to replace the others. Identity controls scope access, sandboxes limit reach, and other controls inspect or govern supported execution paths. Evaluate how the layers work together.

Is a kill switch enough for agent security?

Emergency termination is useful, but evaluate scoped intervention as well. A control switch can constrain or block a proposed action while preserving permitted work. Implementations differ in scope and recovery behavior.

Can Ackuity connect to Microsoft Copilot Studio?

Ackuity supports API integration for Microsoft Copilot Studio. A Kubernetes sidecar is a different connection route. Verify prerequisites and demonstrated action coverage for the workflow you plan to use.

What is the best approach to AI agent security?

A layered approach tested against your workflow: scoped identities, model and tool controls, isolation, monitoring and independent action authorization where supported. No checklist alone proves effectiveness.

What is an agent sandbox?

An isolated execution environment with configured resource and network boundaries. It limits what the workload can reach; evaluate how action-level authorization complements it. Ackuity's OpenShell integration is in development.

Is Ackuity an AI gateway?

Ackuity is the Agent Execution Control Switch. It supports sidecar execution control, hosted-platform API integration and observe-only event pull, rather than relying only on requests routed through an AI gateway.

What is the Goldilocks zone in AI agent security?

Ackuity uses the term for placement beside the agent in a supported execution path: not inside the agent, not at the gateway. The control must remain independent and check the action before execution.

NEXT STEP

Agents are going to act. Decide which actions run.

Show us the actions you need to control, and we’ll show you where Ackuity fits.

Request assessment