AI agent security layers explained
AI agent security splits into five layers: endpoint, browser, cloud workspace, identity and gateway. Each sits in a different place, sees different events and blocks different actions. Cover first the layer where your agents already run, then add the layer that closes the biggest gap, using vendor documents read on 16 Nov 2026.
Facts read on .
What are the five layers of AI agent security?
The five layers are the endpoint, the browser, the cloud workspace, identity and the gateway. They are five places where a control can see an agent act, and no one place sees them all.
OWASP's Excessive Agency entry names three root causes that every layer tries to limit: excessive functionality, excessive permissions and excessive autonomy. Controls at each layer cut one or more of the three.
Treat the layers as a coverage grid rather than a ladder. A control at one layer does not make the next one unnecessary, and a gap at one layer can be closed only at another. A coding agent on a laptop, a browser agent filling a web form and a Copilot Studio agent in a business unit are three different agents with three different control points.
Sources: OWASP LLM06:2025 Excessive Agency.
What does the endpoint layer see?
The endpoint layer sees agents running on laptops and developer machines: desktop assistants, coding agents, command-line agents and local MCP servers. A control here can read the tool call before the agent executes it.
Claude Code's hooks show the mechanism: a PreToolUse hook runs before a tool call executes and can block it. Vendors build on such hooks. Pillar Security says its Runtime Guardian sits in the pre-tool hook of Claude Code, Codex CLI and Cursor, and Operant says its Endpoint Protector inspects AI IDEs, coding agents, desktop and MCP clients.
Local controls have one more advantage: they work when the laptop is off the corporate network. Pillar Security says its endpoint decisions run locally and keep working offline. Ask every endpoint vendor where the decision is made, on the device or in a cloud service, and what happens when the connection drops.
Sources: Claude Code hooks reference; Pillar Security: Agentic Endpoint; Operant AI: Endpoint Protector.
What does the browser layer see?
The browser layer sees AI websites, AI extensions and agents that act inside a browser, such as filling forms or downloading files. Network inspection alone sees the traffic but not the page actions.
Vendors place controls in an extension or a managed browser. Cato says its client covers browser AI and desktop AI apps and can monitor, anonymize or block a prompt, and Island builds agent controls around its enterprise browser, including a headless browser for agents.
Browser controls also cover the human side of AI use. Staff paste text into chat sites, upload files and install AI extensions, and those are the actions a browser control can see and stop. List the AI extensions installed in your fleet before you evaluate any product.
Sources: Cato Networks: AI Endpoint; Island: agentic identity.
What do the cloud workspace layer and identity layer add?
The cloud workspace layer covers agents that makers build inside SaaS and cloud platforms, and the identity layer covers the credentials and permissions every agent carries. Neither runs on a device you manage.
For Copilot Studio, Microsoft documents that agents can be represented as Entra identities and governed with Conditional Access once Agent 365 is onboarded. AWS documents a guardrail setting you attach to a Bedrock agent. Both show that the platform is a control point of its own.
Platform controls differ by platform, so a team using three agent platforms may manage three policy surfaces. A third-party product can unify them with one inventory and one policy language, but check whether it only discovers agents on each platform or can block actions there.
Sources: Microsoft Copilot Studio security and governance; Amazon Bedrock: associate a guardrail with your agent.
What does the gateway layer add?
A gateway adds one control point on the path between agents and models, tools or MCP servers. It sees every call routed through it and sees nothing that bypasses it.
Palo Alto Networks documents an AI Gateway that sits between users or applications and each LLM provider. Cisco publishes a plugin that inspects LLM and MCP requests on a Kong gateway. A gateway fits server-side agents well and fits laptops less well.
Fail-closed behavior matters at a gateway. If the gateway is unreachable, a fail-closed plugin blocks the call and a fail-open plugin lets it through. Cisco documents fail-closed defaults for its checks; confirm the same for any gateway you adopt.
Sources: Palo Alto Networks Prisma AIRS AI Gateway overview; Cisco AI Defense plugin for Kong Gateway.
Which layer should come first?
Start with the layer where your agents run today and where a block is cheapest to deploy. For many companies that is the endpoint, because developers already run coding agents and an MDM script is the install path.
If your agents live in Copilot Studio or Bedrock, start with the cloud workspace and identity layers instead. Add the gateway when agents call external models and tools from servers. Confirm the picks against the Layer Coverage Map before you buy.
Keep a one-page layer map for your own estate: agents by layer, the control at each, the owner and the date last tested. Update it when a new agent platform is approved. The Layer Coverage Map on this site shows what each vendor documents so you can fill the map without starting from blank.
Sources: What is Amazon Bedrock AgentCore; Bay: agent runtime enforcement.
Sources
- OWASP LLM06:2025 Excessive Agency (genai.owasp.org), read
- Claude Code hooks reference (docs.anthropic.com), read
- Pillar Security: Agentic Endpoint (pillar.security), read
- Operant AI: Endpoint Protector (operant.ai), read
- Cato Networks: AI Endpoint (catonetworks.com), read
- Island: agentic identity (island.io), read
- Microsoft Copilot Studio security and governance (learn.microsoft.com), read
- Amazon Bedrock: associate a guardrail with your agent (docs.aws.amazon.com), read
- Palo Alto Networks Prisma AIRS AI Gateway overview (docs.paloaltonetworks.com), read
- Cisco AI Defense plugin for Kong Gateway (github.com), read
- What is Amazon Bedrock AgentCore (docs.aws.amazon.com), read
- Bay: agent runtime enforcement (bay.io), read