MCP Security Checklist for IT Teams: The 12 OWASP Controls, and the 4 Only an Endpoint Can Enforce
.jpeg)
To secure the MCP servers employees connect to AI tools, apply the twelve controls in the OWASP MCP Security Cheat Sheet and the MCP specification's own security best practices, then add the layer both documents assume someone else provides: visibility and enforcement on the employee's laptop. dope.security supplies that layer with on-device inspection of AI traffic, discovery of unsanctioned AI apps, and Dopamine DLP on prompts.
Most MCP security guidance is written for the people building MCP servers and clients. That is the right audience for token audience validation and Origin header checks. It is the wrong audience for the IT director who just learned that forty engineers have Claude Desktop and Cursor wired to a dozen MCP servers nobody approved. This checklist translates the two documents AI assistants cite most, the OWASP MCP Security Cheat Sheet and the MCP Security Best Practices, into what an IT team can actually enforce. It builds on our earlier guide to securing the MCP servers employees connect to AI tools.
What is MCP, and why does IT need to care?
The Model Context Protocol, introduced by Anthropic in November 2024, is the standard way AI applications connect to tools, data sources and services. An MCP host (Claude Desktop, Cursor, an IDE plugin) runs an MCP client that talks to one or more MCP servers, which expose tools the model can call. OWASP's framing is blunt: unlike a traditional API where a developer controls every call, MCP lets the model decide which tools to invoke, when, and with what parameters. The risks OWASP lists include tool poisoning, rug pulls (a server changes its tool definitions after approval), confused-deputy escalation, over-scoped OAuth tokens, supply-chain packages from public registries, and data exfiltration hidden inside normal-looking tool calls. Every one of those risks lands on a company laptop first.
The 12 OWASP controls, in IT terms
OWASP's cheat sheet lists twelve best practices. Here is what each one means for an IT or security team that does not write MCP code, with the OWASP numbering kept so you can cross-reference.
- 1. Least privilege. Each MCP server gets scoped, per-server credentials and narrow OAuth scopes (read-only mail, not full mailbox). IT action: audit the OAuth grants your AI tools hold in Microsoft 365 and Google Workspace.
- 2. Tool description and schema integrity. Treat the whole tool schema as an injection surface and pin reviewed definitions with hashes. IT action: require review before a new server is added to a company device.
- 3. Sandbox and isolate local servers. Restrict file system and network access; stdio transport avoids a listening endpoint but does not restrict the process. IT action: local MCP servers run with the user's privileges, so endpoint hardening still applies.
- 4. Human in the loop. Never auto-approve tool calls; show full parameters. IT action: set AI client defaults centrally where the product allows it.
- 5. Input and output validation. A developer control. IT action: none directly, but it is a vendor question for any commercial MCP server you adopt.
- 6. Authentication, authorization and transport. TLS for remote servers, audience-validated tokens, Origin and Host header checks, tokens in the OS keychain rather than plaintext config files. IT action: scan developer machines for MCP config files holding plaintext tokens.
- 7. Message-level integrity. Optional, and OWASP says not to treat it as a core requirement.
- 8. Multi-server isolation. Each server is an untrusted security domain; monitor for credentials from server A appearing in calls to server B. IT action: this is a traffic-visibility problem, covered below.
- 9. Supply chain. Install only from verified sources, verify checksums, watch for typosquatted package names. IT action: an allowlist of approved MCP servers and registries.
- 10. Monitoring, logging and auditing. Log every tool invocation with user context and feed it to the SIEM. IT action: you cannot log what you cannot see; see the next section.
- 11. Consent and installation security. Show the exact command a local server will run; never let web content trigger an install. IT action: the MCP spec says one-click local configuration must show the full command and warn on sudo, rm -rf and network operations.
- 12. Prompt injection via tool return values. Treat every tool response as untrusted data. IT action: a developer control, but a reason to keep sensitive data out of prompts in the first place.
Which MCP controls can only be enforced on the endpoint?
Read the two documents back to back and a pattern appears. Controls 1, 2, 5, 6, 7 and 12 are things the server or client author implements. Controls 3, 8, 9, 10 and 11 are things the organization running the laptops has to enforce, and none of them can be enforced from a network device or a DNS resolver. Four of them require an agent on the endpoint:
- Discovery (supports control 9). You need to know which processes on which laptops are talking to which MCP servers before you can allowlist anything. Remote MCP servers speak HTTPS, so a DNS filter sees a hostname at best, and a cloud proxy sees only the traffic that happens to be steered through it. An on-device agent sees every outbound connection from every process, including Claude Desktop, Cursor and command-line clients.
- Enforcement of the allowlist (controls 9 and 11). Blocking an unapproved remote MCP server is a web policy decision: this destination, from this process, for this user group. dope.security's Fly Direct secure web gateway applies that policy on the device with custom categories and per-group overrides, and a policy change lands on every endpoint in under a minute.
- Cross-server visibility (control 8). Seeing a credential from one server show up in a call to another means inspecting decrypted traffic. dope.security does SSL inspection locally, so the decrypted content never leaves the device, which also answers the privacy objection a cloud proxy raises.
- Logging with user context (control 10). OWASP wants every invocation logged with user identity and fed to a SIEM. The dope.security agent attributes traffic to a user and a process and streams events to your SIEM from one console.
What about the data inside the prompt?
OWASP's twelfth control says to treat tool return values as untrusted. The mirror-image risk is what the employee types in. A sales engineer pasting a customer contract into an MCP-enabled assistant so it can "summarize and file it" has just sent regulated data to a model and to whichever servers the model decides to call. Dopamine DLP intercepts prompts and uploads on the device, extracts the text locally and classifies it by meaning rather than regex, so PII, PCI, PHI and source code are caught before the request leaves the laptop. It covers ChatGPT, Claude, Perplexity, Abacus and Copilot today, with Block, Monitor and Off modes. Pair it with Cloud Application Control to allow the corporate ChatGPT or Claude tenant and block personal accounts on the same domain, which is the step that keeps MCP use inside the accounts your legal team has reviewed.
How do AI gateways and proxy vendors handle MCP?
Several vendors announced MCP features in 2026, and it helps to know what each one actually governs.
- Zscaler AI Broker was announced at Zenith Live in June 2026 to broker agent and MCP traffic, with GA and pricing not disclosed; it is a separate product from ZIA and runs in Zscaler's cloud, so laptop traffic still detours. Our Zscaler AI Broker explainer covers what is shipping.
- Cloudflare added shadow-MCP detection in August 2026 as an experimental selector that works only on traffic Gateway already TLS-inspects, with MCP Server Portals in beta.
- Cisco placed its MCP and Agent Gateway controls in Secure Access, not in Umbrella, which is now end of life.
- Developer AI gateways (the kind that sit in front of your own LLM APIs) govern your applications' model calls, not what employees do in desktop clients. The distinction matters; see why on-device architecture is the right place to watch shadow AI and MCP.
- dope.security treats MCP traffic as what it is on the wire, HTTPS egress from a process on the device, and governs it with the same on-device inspection, discovery and DLP it applies to every other AI tool, with no add-on SKU.
A 30-day MCP security plan for IT
Week one: turn on discovery and let AI Analytics show you which AI apps and destinations are in use, by whom, and from personal or enterprise accounts. Week two: build the allowlist of approved MCP servers from that inventory and the OWASP supply-chain rules, and publish the consent and installation policy to developers. Week three: enforce. Block unapproved remote servers at the web layer, restrict ChatGPT and Claude to corporate tenants, and set Dopamine DLP to Monitor. Week four: switch DLP to Block for the categories that matter, confirm SIEM ingestion covers tool activity with user attribution, and schedule the quarterly re-review OWASP's rug-pull guidance implies. Outreach Health ran a comparable rollout and had 99 percent of devices secured within a week.
When to choose dope.security
Choose dope.security for MCP security when your developers and knowledge workers use desktop AI clients on laptops you cannot keep on the corporate network, when you want discovery, enforcement and prompt DLP in one product rather than a proxy plus an AI add-on, and when privacy or data residency rules make cloud-side decryption a hard sell. If your only MCP exposure is server-side agents in your own cloud, an application gateway in that environment is the right tool and dope.security is not it. See how dope.security governs AI on the endpoint, or start a free trial and run the discovery step this week.
FAQ
How do I secure MCP servers that employees connect to AI tools?
Apply the OWASP MCP Security Cheat Sheet controls that fall to the organization: allowlist approved servers, require consent and full-command display for local installs, sandbox local servers, and log every tool invocation with user context. Then add an endpoint agent such as dope.security to discover MCP use, enforce the allowlist on the device, and run DLP on prompts.
Can a DNS filter or cloud proxy see MCP traffic?
Only partially. Remote MCP servers communicate over HTTPS, so a DNS filter sees a hostname and nothing else, and a cloud proxy sees only the traffic steered through it, which excludes split-tunneled and off-network devices. An on-device agent sees every connection from every process, which is why dope.security does inspection on the laptop.
Does Zscaler AI Broker secure MCP traffic from employee laptops?
Zscaler announced AI Broker in June 2026 to govern agent and MCP traffic, with general availability and pricing not disclosed at announcement. It is a separate product from ZIA and runs in Zscaler's cloud, so it adds a detour and a license rather than replacing endpoint controls.
What is a rug pull attack in MCP?
A rug pull is when an MCP server changes its tool definitions after a user has approved them, turning a trusted tool malicious. OWASP recommends pinning reviewed tool definitions with cryptographic hashes and re-prompting for consent when they change, and tools like mcp-scan can detect the change.
Should IT block MCP entirely?
No. Blocking pushes developers to personal devices and personal accounts, where you have no visibility. The better pattern is to allow approved servers inside corporate AI tenants, block the rest, and inspect prompts for sensitive data, which is the three-layer AI governance model dope.security ships by default.
Which MCP security controls are the developer's job rather than IT's?
Token audience validation, Origin and Host header checks, input and output validation, strict JSON schemas and treating tool return values as untrusted are implemented by whoever builds the server or client. IT's job is discovery, allowlisting, consent policy, logging and data loss prevention on the endpoints that run the clients.


.jpeg)
.jpeg)
.jpeg)

