AI agent security at the endpoint layer
Endpoint coverage means a vendor documents controls for desktop agents, coding agents, command-line agents and local MCP servers on employee machines. On 16 Nov 2026, 13 of 16 vendors had a yes at the endpoint layer, 3 had partial and 0 had none found. Levels are from vendor documentation, not lab tests.
Facts read on .
Endpoint: every vendor
| Vendor | Level | Note | Evidence |
|---|---|---|---|
| Aim Security | Yes | Local agents are found and tool calls blocked on the endpoint. | aim-security-11, aim-security-40, aim-security-41 |
| Bay | Yes | Endpoint product covers coding agents, desktop assistants and MCP servers. | bay-14, bay-72, bay-73 |
| Cisco | Yes | DefenseClaw hooks for named coding agents on Linux, macOS and Windows. | cisco-10, cisco-11, cisco-12 |
| CrowdStrike | Yes | Falcon sensor visibility plus agent-level hooks for Claude Code, Codex, Copilot, OpenCode and Cursor on Windows, macOS and Linux. | crowdstrike-02, crowdstrike-09, crowdstrike-10 |
| Island | Yes | Island Desktop inventories and governs agents, MCP servers and skills on managed devices. | island-04, island-05 |
| Lasso Security | Yes | Lasso says it inventories Claude Code, Cursor and Windsurf connections and governs endpoint AI agentlessly. | lasso-security-19, lasso-security-26, lasso-security-12 |
| Microsoft | Yes | Defender for Endpoint discovers and protects local agents. | microsoft-11, microsoft-19 |
| Noma Security | Yes | Endpoint discovery via EDR or MDM and named coding agents. | noma-security-08, noma-security-09 |
| Operant AI | Yes | Endpoint Protector inspects coding agents, MCP clients and plugins on macOS, Windows, Linux. | operant-ai-09, operant-ai-10 |
| Pillar Security | Yes | Coding agent discovery, posture and pre-action blocking on managed machines. | pillar-security-05, pillar-security-09 |
| SentinelOne (Prompt Security) | Yes | Dedicated AI endpoint agent that detects new agents and enforces runtime policy. | sentinelone-13, sentinelone-32 |
| Straiker | Yes | Coding agents such as Cursor, Copilot, Claude Code and Windsurf covered via a thin client; endpoint discovery stated. | straiker-06, straiker-07 |
| Zenity | Yes | Coding and personal agents on laptops and desktops are discovered and blocked via native hooks and MCP gateway. | zenity-19, zenity-50 |
| Check Point | Partial | Desktop agent and IDE coverage is claimed under Workforce AI Security; OS list and local control detail none found. | check-point-08, check-point-10, check-point-11 |
| Palo Alto Networks | Partial | Only community example hooks for coding assistants in a best-effort GitHub repo; endpoint governance is in a separate profile. | palo-alto-networks-04, palo-alto-networks-50 |
| WitnessAI | Partial | Sees developer-machine agents through network traffic and Cursor, Claude Code and Codex workflows; no endpoint agent documented. | witnessai-25, witnessai-26 |
What agents run at the endpoint layer
Coding agents, desktop assistants, command-line agents and locally installed MCP servers run as processes on employee laptops. Cisco says an interactive coding agent can read a repository, run the build and push a change, and acts with the full access rights of the developer who started it. Noma says such agents, MCP servers and skills carry the same permissions and credentials as the people who installed them.
Sources: Protect enterprise coding agents with Cisco DefenseClaw and NVIDIA OpenShell.
What a control at the endpoint layer can and cannot see
A control on the machine can see local processes, commands, configuration files, skills directories and file access, and it can act before a local action completes. Noma says endpoint activity leaves local evidence in configuration files, skills directories, connector history and running processes.
It cannot see agents that run in a vendor's cloud. Operant notes that very little Copilot activity has a local process to instrument, because the prompt is answered with content pulled from cloud services the user never opened on the machine.
Sources: Noma Extends Agent Security and Governance to the Employee Endpoint; Operant: coverage for Microsoft Copilot and embedded AI assistants.
Who documents the endpoint layer
Yes on 16 Nov 2026: Aim Security, Bay, Cisco, CrowdStrike, Island, Lasso Security, Microsoft, Noma Security, Operant AI, Pillar Security, SentinelOne (Prompt Security), Straiker, Zenity.
Partial: Check Point, Palo Alto Networks, WitnessAI.
None found: none.