FAQ, read 12 Nov 2026
30 AI endpoint security questions, answered with names and numbers
- What does AI endpoint security cover?
- It covers AI agents, assistants, MCP servers and browser extensions that run on employee laptops, in browsers and in developer tools. The job has two parts: show which of them exist and what they do, then stop the ones that act outside policy. This site maps six surfaces: desktop AI apps, IDE assistants, command-line agents, MCP servers, browser extensions and SaaS copilots.
- How is AI endpoint security different from EDR?
- EDR records and responds to process, file and network behavior on a device. It generally cannot tell a shell command run by a coding agent from one typed by a developer, and it does not read prompts or tool calls. AI endpoint security adds agent-level context: which tool, which session, which tool call. Some products build on EDR telemetry.
- How is it different from DLP?
- DLP inspects content moving to a destination and was built around user actions such as paste, upload and copy. Microsoft's endpoint DLP, for example, has rules for paste to supported browsers and upload to restricted domains. Agents add reads and tool calls with no user action, often from a terminal. AI endpoint security aims at those actions, not only at the destination.
- How is it different from CASB?
- A CASB watches SaaS use through APIs, proxies or logs. It sees cloud apps and OAuth-connected services, not what runs on the laptop. A local MCP server or a command-line agent using a personal API key never touches the SaaS API. CASB data is useful for the SaaS copilot surface and complements endpoint visibility.
- Which agent surfaces matter most?
- It depends on who works at your company. For engineering teams, command-line agents, IDE assistants and local MCP servers carry the highest-impact actions: shell commands, credential access and package installs. For knowledge workers, browser extensions, desktop AI apps and SaaS copilots hold more of the data-paste and oversharing risk. Most companies need to cover all six eventually.
- What actions should a tool be able to see?
- Six matter: file read, shell command, network call, credential access, data paste and package install. A product should say, for each surface, which of these it sees, which it can stop, and which component on the device does so. A product that lists surfaces without listing actions has not told you what it covers.
- What does blocking before completion mean?
- It means the product decides whether an action may proceed before the action finishes: the command does not run, the file is not read, the paste does not land. Alerting after the fact is detection. Both are useful, but they are different. Ask a vendor to demonstrate a block on your own machine, and to show what the user sees.
- Why is the difference between detection and prevention important here?
- An agent can read a credential file or run a package install in seconds, and an alert that arrives an hour later cannot recall what was sent. OWASP names excessive functionality, permissions and autonomy as root causes of agent harm, which a pre-action check can limit. Detection still matters for inventory and investigation.
- What does agentless mean for AI endpoint security?
- It usually means no new sensor ships to the laptop. The data may come from an MDM profile, a browser extension, an OS API, existing EDR telemetry or a SaaS API connector. Each sees a different slice. Browser extensions miss terminal agents, and API connectors miss processes. Ask which component produces the data for each surface.
- Can MDM alone manage AI tools on laptops?
- MDM can push settings to tools that read managed policy. Anthropic documents that Claude Code accepts endpoint-managed settings through macOS managed preferences, Windows registry or a managed file. Intune can apply custom macOS profiles. But MDM is a configuration channel, not a sensor: it does not show what the tools do or find tools you did not know about.
- Where do local MCP servers live on a laptop?
- They are defined in each client's config file. Claude Desktop uses claude_desktop_config.json (under Application Support on macOS, APPDATA on Windows, .config on Linux). Cursor uses .cursor/mcp.json in a project and ~/.cursor/mcp.json globally. Claude Code uses .mcp.json in a project and ~/.claude.json for user and local scope.
- Why are local MCP servers a risk?
- A stdio MCP server is a program the client launches with the user's privileges. The MCP security best practices page lists attacks including a malicious startup command in a client configuration and a malicious payload inside the server, and asks for exact-command consent and sandboxing with restricted file system and network access.
- How do I control MCP servers on laptops?
- Control is per client. Claude Code reads a managed-mcp.json from a fixed system path for exclusive control, plus allow and deny lists. VS Code has a ChatMCP access policy and an allowed-servers list. For clients without an equivalent policy, none found on 12 Nov 2026 in the docs read, so monitor the config files and block launch commands.
- Does a project folder's MCP config add risk?
- Yes. Cursor and Claude Code both read an MCP config from inside the project (.cursor/mcp.json, .mcp.json), and project files can be committed to source control. A cloned repository can propose a server to a developer's machine. Claude Code has project approval and workspace trust steps for this, and developers need to read those prompts.
- What is the difference between stdio and remote MCP servers?
- With stdio, the client launches the server as a subprocess on the same machine and talks to it over standard input and output. A remote server is reached over Streamable HTTP. The MCP specification defines both transports. Stdio servers are the endpoint concern, because they run locally with the user's rights.
- What risks do coding agents bring to developer machines?
- Coding agents run shell commands, read the project and the home directory, install packages and call network endpoints, all as the developer. Developer laptops often hold source code, signing keys and cloud credentials. The main risks are excessive permissions, unreviewed MCP servers and credential exposure in project files. Managed settings and sandboxing reduce them.
- Can I enforce settings on Claude Code across a fleet?
- Anthropic says Claude Code supports endpoint-managed settings deployed through native OS policies (macOS managed preferences, Windows registry) or managed settings files, and server-managed settings fetched from Anthropic's servers for organizations without MDM. Endpoint-managed settings are the stronger route where you run MDM.
- What does VS Code offer for enterprise control of agents?
- Microsoft publishes VS Code enterprise policies including ChatMCP (access to installed MCP servers), ChatAllowedMcpServers (an allowlist, with deny rules taking precedence), and policies for agent hooks and plugins. These are deployed through your device management tooling. They govern the editor's features, not other tools on the machine.
- What are browser agents, and why do they need separate controls?
- Browser agents are AI features that read pages and act on them: click, fill forms and move across tabs. Examples include Perplexity's Comet. They act inside the user's signed-in sessions. Anthropic's November 2025 post says no browser agent is immune to prompt injection. Controls are extension policy, site scoping and confirmation steps.
- How do AI browser extensions get their access?
- Through the permissions in their manifest. Google's documentation explains that permission warnings describe API capabilities and that a broad host permission lets an extension act on every site. An AI extension that reads pages usually requests that breadth. Review permissions, publisher and update history before allowing it, and restrict by policy.
- Can I control which extensions run in Chrome and Edge?
- Yes. Both support the ExtensionSettings policy, which can control update URLs and blocked permissions, and sits alongside install allowlists, blocklists and force-install lists. Microsoft documents it for Edge. The strongest setup is default deny with a reviewed allowlist by extension ID. Also manage Incognito so extensions are not bypassed.
- What does AI DLP on the endpoint do?
- It applies data loss rules to AI-related actions: pasting text into an AI site, uploading a file, or copying from protected files. Purview endpoint DLP, for example, evaluates pasted content in supported browsers and can warn, block or audit. It covers people at keyboards well. Agent reads and API calls from terminals are the thinner area.
- Does Microsoft Purview block pasting into ChatGPT?
- Microsoft's documentation says Purview DLP policies can warn or block users from sharing sensitive information with third-party generative AI sites accessed through a browser, with pasting credit card numbers into ChatGPT as its example. For Chrome, the Purview extension must be installed. Check the current platform matrix, since some entries are marked preview.
- How do I detect shadow AI on endpoints?
- Combine signals: process and binary inventory, browser extension inventory, network and DNS records, agent and MCP config files, and OAuth grants. Each sees a slice. DNS shows a domain but not a prompt. OAuth grants miss personal accounts. A tool installed on purpose in a test run shows which signals work.
- Which OAuth grants should I review for AI apps?
- Start with apps holding write or full-mailbox scopes, then any app consented by several users in a short period. Microsoft documents how to review the permissions granted to applications in a Microsoft Entra tenant, including apps added through user or admin consent. This catches apps on unmanaged devices but not personal accounts or API keys.
- What should I ask an AI endpoint security vendor?
- Ask which surfaces and which actions it covers, which component on the device produces the data, whether it can block before an action completes, what a user can do to disable it, and how it handles a tool launched from a terminal. Ask for a trial where you run one tool per surface and compare its log with what you did.
- How do I test a vendor claim?
- Run one tool per surface on a test machine: a desktop app, an IDE assistant, a command-line agent, a local MCP server, a browser extension and a SaaS copilot sign-in. Make each do something harmless and observable. Compare the product's record with your actions, then ask to see a block. Score what you saw, not the data sheet.
- How does this site score vendors?
- Seven criteria with fixed weights set before any evidence was scored: sees 25, stops 25, deployment 15, beyond the endpoint 10, policy and governance 10, transparency 10 and proof 5. Each criterion is scored from sourced evidence. The weights are the same for every vendor.
- What happens to a vendor claim that nobody else confirms?
- A vendor's own claim is capped at 7 of 10 on a criterion until a second source confirms it, such as a write-up from another publisher or a public artefact like documentation, code or a release record. The claim is written as the vendor's. Once confirmed, it can score higher. This keeps marketing copy from scoring the same as verified behavior.
- What score does a missing feature get?
- A feature with no evidence found scores 0 on that criterion. The page records it as none found on a date, for example none found on 12 Nov 2026, rather than as a statement that the vendor lacks it. If the vendor later publishes evidence, the score is revisited and the date updates.
Where the answers come from
Sources
- Anthropic: Claude Code docs, security, read 12 Nov 2026
- Anthropic: Claude Code docs, permissions, read 12 Nov 2026
- Cursor: security, read 12 Nov 2026
- Cursor: docs, agent security, read 12 Nov 2026
- Microsoft: VS Code docs, Copilot security, read 12 Nov 2026
- Microsoft: VS Code docs, agent tools, read 12 Nov 2026
- OpenAI: Codex docs, agent approvals and security, read 12 Nov 2026
- OpenAI: Codex docs, configuration reference, read 12 Nov 2026
- raw.githubusercontent.com Google, read 12 Nov 2026
- raw.githubusercontent.com Google, read 12 Nov 2026
- Anthropic: Claude Code docs, sandboxing, read 12 Nov 2026
- Microsoft Learn: Intune custom settings for macOS, read 12 Nov 2026
- Anthropic: Claude Code server-managed settings, read 12 Nov 2026
- modelcontextprotocol.io: security best practices, read 12 Nov 2026
- modelcontextprotocol.io: transports specification 2025-11-25, read 12 Nov 2026
- Perplexity: Comet, read 12 Nov 2026
Last reviewed 12 Nov 2026.