AI Agent Security Platforms: Where Sensors, Gateways, Hooks and Proxies Each Start (and What They Miss)
.jpg)
Last updated: November 20, 2026
Short answer: An AI agent security platform enforces policy from one of three places: on the device (a sensor or browser extension), inside the agent (hooks, an SDK or an API), or in the network path (a proxy or gateway). Independent analyst Andrew Green's new map shows most vendors starting in one place and expanding into the others, and where a vendor starts predicts what it can't see.
That second point is the useful one for anyone comparing agentic AI security vendors. Labels like "agent security" now cover very different products. This post walks through Green's map, explains what each architecture sees and misses, and shows where dope.security sits on it. We make dope.security, so read that section knowing who wrote it.
What is an AI agent security platform?
An AI agent security platform is a product that gives security teams visibility and control over AI agents and the tools they use, by collecting data or enforcing policy at a specific point: the device, the agent's own code, or the network path. Which point it starts from determines which agent actions it can observe and which it can't.
That definition is deliberately about placement, not features. Feature lists converge quickly. Placement doesn't change as easily, and it sets hard limits on what a product can know.
How does Andrew Green map AI agent security vendors?
On November 20, 2026, Andrew Green published "Where Agent Security vendors start, and where they all end up" on his Substack, Green Abstracts, which describes itself as non-sponsored analysis of enterprise IT markets. His premise: "you can use product architecture to classify and extrapolate what they can and cannot do." His complication: nearly every vendor is now adding hooks, gateways, sensors and browser extensions, whatever it started with.

How to read the chart
In Green's words, "A box on the chart is a vendor's first or primary collection/enforcement point." The colors group the boxes:
- Blue, on the device: sensor, browser or extension. "Things installed on the user's machine, independent of any one agent."
- Yellow, in the agent code: API, SDK, hook. "Things the agent, its harness or its developer has to call."
- Magenta, in the network path: proxy, gateway. "Things the agent's traffic passes through."
Arrows show where each category is typically expanding. The chart is a map of starting points, not a ranking.
What does each AI agent security architecture see and miss?
Each row below summarizes Green's description of the category, not our assessment of any vendor in it.
| Architecture | What it sees | What it misses |
|---|---|---|
| Gateway | Every call that crosses the network, tied to an identity, for any harness | "anything that does not cross the network: shell commands, local MCP servers, local models" |
| Catalogue (supply chain) | Which MCP servers, packages and skills are approved | Runtime behavior; Green calls it "the least differentiating part of any stack" |
| Endpoint sensor | OS-level telemetry: processes, files, connections and their lineage, "independent of what the agent says" | The agent's stated intent |
| Hooks | "the prompt and the pending action together" | Anything the agent doesn't report; hooks "live in configuration the user or the agent can change" |
| In-process | The database query, API call and file behind each action | Agents you don't build and own |
| Verdict API | An allow or block decision on a prompt or tool call | The enforcement point, which a partner controls |
| In-line proxy | Traffic that passes through the proxy, "agnostic to the runtime environment" | What never reaches the proxy |
Endpoint sensors and browser extensions
A sensor sees what actually ran on the host, whatever the agent claims, and Green notes it "can block a process outright." Browser extensions cover chat apps, web agents and agentic browsers. The tradeoff is that a sensor sees actions, not intent.
Hooks, SDKs and APIs
Hooks are attractive because they're "the only place that sees the prompt and the pending action together." Green's caution is that "they only see what the agent reports," and the agent or user can change them. That's why many hook-first vendors are adding sensors or gateways.
Gateways and proxies
Anything in the network path shares one boundary. As Green puts it for gateways, what it "never sees is anything that does not cross the network." A gateway in front of MCP traffic is simple to deploy and identity-aware, but a shell command that stays on the laptop is invisible to it. For a deeper look at that category, see what an AI gateway controls.
Catalogues, in-process and verdict APIs
Catalogues vet what gets installed. In-process tools see the most detail per action but only for agents you build. Verdict APIs, which Green calls "first-wave products," sell the decision while someone else's gateway or SDK enforces it.
Why are AI agent security vendors converging?
Green's central observation is that vendors are ending up in the same place. Gateway vendors add device components. Hook vendors add gateways and sensors. Sensor, hook and browser-extension bundles show up "quite often." He also notes that "cloud proxies are moving onto the device," for one reason: "shell commands and local MCP servers never reach the network."
He describes proxy vendors differently: "These seem the most stable approaches, i.e. these vendors don't see a need to deploy other mechanisms. They are agnostic to the runtime environment and operate at a lower level in the stack." That's an observation about the category's behavior, not a rating of any vendor.
Our take: as agentic AI security companies converge, the label on a product tells you less every quarter. A better test is the question you need answered. "Which agents are running on our laptops and what are they connecting to?" calls for a different control than "what did this agent decide to do, and why?"
Where does dope.security sit on Green's map?
Green's chart places dope.security in the in-line proxy box, alongside other proxy vendors. The detail worth adding is where dope.security's proxy runs.
Green notes that cloud proxies are moving onto the device. dope.security's proxy runs on the device by design, an architecture it calls Fly Direct. Put plainly, dope.security is a secure web gateway whose proxy runs on the managed Mac or Windows laptop, not in a cloud data center.
That placement matters for AI agents in specific ways:
- Non-browser traffic. Interception happens at the OS level (a Windows WFP driver, a Mac network extension), so traffic from non-browser apps is intercepted as well as browser traffic, apart from apps and domains on the bypass lists.
- AI app and MCP visibility. dope.security sees web traffic from AI apps and agent tools on the device, and its MCP Usage report tracks which MCP servers are used across the organization. The MCP Usage report shows which MCP servers your organization uses, by whom, and how much data flows through them. It's a visibility report and doesn't show individual tool calls. More detail is in our posts on shadow AI agents and MCP servers and MCP governance.
- Account controls. Cloud App Controls let admins restrict supported apps to company accounts, including ChatGPT (Enterprise), Claude (Team and Enterprise) and GitHub (Enterprise). Cloud App Controls also support a warn mode for Gemini and Copilot, an access warning rather than a DLP mode.
- Data in prompts and uploads. Dopamine DLP can monitor or block sensitive uploads and prompts to a documented list of AI apps, including Microsoft Copilot (browser), Gemini, NotebookLM, Perplexity (browser) and Abacus AI (browser).
- Investigation. On endpoints running dope.security, SIEM logs include the process name and the process call tree (the parent-child chain) for the requests they record.
What dope.security doesn't cover
Green's network-path limit applies here too. dope.security's proxy sees network traffic from agents and tools on the laptop, including non-browser apps, but actions that never cross the network, such as a local model or a shell command that makes no network request, are outside its view.
It's also a different kind of product from the hook and in-process vendors on the map. dope.security does not provide agent hooks, agent runtime protection or tool-call inspection, and does not cover servers, Linux hosts or CI runners; it protects managed Mac and Windows endpoints.
So it's one layer. Teams building their own agents, or running agents on servers and in CI, will pair it with hooks, sensors or in-process controls for those environments. For how device-level control differs from a gateway, see endpoint AI governance.
How should you evaluate an AI agent security platform?
Ask every vendor the same questions, whatever its label:
- Where does it enforce first? Device, agent code or network path?
- What happens to actions that never cross the network? Shell commands, local models and local tools.
- Which environments does it cover? Managed laptops, servers, CI runners, SaaS-hosted agents.
- Can the agent or the user change the control? Configuration that lives with the agent can drift.
- What does it see at the tool-call level? Some products see each action; others see only connections.
- What's logged, and where does it go? Can you trace a request back to a process and a user?
- What does it add when it expands? A new mechanism, or the same one redeployed somewhere else?
The answers usually point to a combination. Green's map is a good way to see which combination you're buying. For a related comparison, see AI firewall vs AI gateway vs endpoint AI governance.
Frequently asked questions
What is an AI agent security platform?
An AI agent security platform gives security teams visibility and control over AI agents and their tools by collecting data or enforcing policy at one point: on the device, inside the agent's code, or in the network path. That starting point determines which agent actions the platform can see and which it misses.
What is agentic AI security?
Agentic AI security is the practice of protecting AI agents that take actions, such as calling tools, running commands and moving data, rather than only answering prompts. It covers which agents and tools are in use, what they can reach and what they actually do, using controls on the device, in the agent's code or in the network path.
What are the main AI agent security architectures?
Andrew Green's map groups them by first enforcement point. On the device: endpoint sensors, EDR piggyback and browser extensions. In the agent code: hooks, SDKs, verdict APIs, in-process tools and supply-chain catalogues. In the network path: in-line proxies, gateways and cloud proxies. Most vendors start in one and expand.
Are gateways enough to secure AI agents?
Not on their own. A gateway sees every call that crosses the network and can tie it to an identity, but as Andrew Green notes, it never sees anything that doesn't cross the network, such as shell commands, local MCP servers and local models. Most teams pair it with device or in-agent controls.
What's the difference between hooks and sensors for AI agents?
Hooks run inside the agent's workflow and see the prompt and the pending action together, but only what the agent reports, and the agent or user can change them. Sensors run on the host and see what actually happened, such as processes, files and connections, regardless of what the agent claims.
Where does dope.security fit among AI agent security vendors?
Andrew Green's chart places dope.security in the in-line proxy box. dope.security is a secure web gateway whose proxy runs on the managed Mac or Windows laptop, not in a cloud data center. It does not provide agent hooks, agent runtime protection or tool-call inspection, and does not cover servers, Linux hosts or CI runners.
Can dope.security see which MCP servers employees use?
Yes. The MCP Usage report shows which MCP servers your organization uses, by whom, and how much data flows through them. It's a visibility report and doesn't show individual tool calls. dope.security also sees web traffic from AI apps and agent tools on the device.
How do I evaluate AI agent security vendors?
Ask where each product enforces first, what happens to actions that never cross the network, which environments it covers (laptops, servers, CI), whether the agent or user can change the control, what it sees at the tool-call level, and what it logs. The answers usually point to a combination of controls.
See what your AI agents already connect to
Green's map is a reminder that no single placement sees everything. The practical first step is knowing which AI apps, agent tools and MCP servers your team already uses, then choosing the controls that cover the gaps.
Start a free trial with your Microsoft 365 or Google sign-in, no payment method needed. Or email sales@dope.security.
Sources
- Andrew Green, "Where Agent Security vendors start, and where they all end up", Green Abstracts, Nov 20, 2026
- Andrew Green, LinkedIn post, Nov 20, 2026
- dope.security, dope.swg and pricing


.jpeg)
.jpg)

