Every time you connect an AI agent to your email, your CRM or a database through MCP (Model Context Protocol, the open standard that Claude, ChatGPT and Gemini use to talk to your systems), you’re trusting a server that almost nobody has actually audited. Enkrypt AI scanned 25,000 MCP servers and 268,000 connected tools and found that 73% had at least one exploitable vulnerability — 143,000 flaws in total. Astrix Security audited 5,205 open-source MCP servers in October 2025 and found that only 8.5% use OAuth; the rest connect with keys that never expire, often stored in plain text.

MCP solves a real problem: before it existed, connecting an AI agent to each tool — Gmail, Slack, a CRM, a database — meant writing a custom integration for every single one. MCP standardizes that connection. The problem is that adoption has moved far faster than governance: anyone can stand up an MCP server in an afternoon, and most companies connect their agents to one without asking who maintains it, what it actually does, or what it can touch.

What an MCP Gateway is, and why your company already depends on one

An MCP Gateway sits between your AI agents and the MCP servers they connect to: it decides which server can talk to which agent, which tools it can call, with what credentials, and what gets logged. Without that layer, every agent connects directly to every MCP server, and access control depends entirely on whatever default configuration that particular server shipped with — which, based on the numbers below, is almost never the right one.

The difference from a classic API integration is this: a traditional API gets installed and reviewed by your IT team, with a contract and a security questionnaire behind it. An MCP server, in most real deployments, gets discovered by a developer in a public repository, tried out, works, and stays connected to production agents. Nobody formally onboarded it as a security asset, because it never went through that process.

The numbers that never show up in the “connect your AI to everything” demo

The demo shows how fast an agent solves a task by reaching into ten different tools at once. What it doesn’t show is whether anyone audited those ten connections. Every figure below comes from an independent 2025 or 2026 study, and none of them are reassuring:

  • 73% of the 25,000 MCP servers Enkrypt AI scanned had at least one real vulnerability — 143,000 flaws in total, including command injection, authorization bypass and prompt injection.
  • In an earlier Enkrypt AI scan of 1,000 servers, 33% had critical vulnerabilities, averaging 5.2 flaws per server. One widely-used MCP server for managing Kubernetes clusters had 26 vulnerabilities — six of them command-injection flaws scored at the maximum severity (CVSS 9.8) — capable of handing over an entire cluster.
  • 53% of the 5,205 servers Astrix Security audited rely only on static API keys that never expire; 79% of those keys are stored in plain environment variables. Only 8.5% use OAuth, the one standard that actually lets you revoke access without touching the server itself.
  • Between January and April 2026 alone, more than 40 distinct CVEs were disclosed against MCP implementations across Python, TypeScript, Java and Rust — issues in the protocol’s foundations, not isolated bugs from one vendor.

The warning that already became real: CVE-2025-6514

This isn’t a theoretical risk. In July 2025, JFrog’s security research team found that mcp-remote — a package that lets AI clients connect to remote MCP servers, downloaded more than 437,000 times — had a command-injection flaw (CVE-2025-6514, severity 9.6 out of 10). A malicious MCP server could execute code directly on the machine of whoever connected to it, with no action required beyond establishing the connection. It was, according to JFrog, the first real-world demonstration of full remote code execution through an MCP connection — not a lab exercise. The flaw affected versions 0.0.5 through 0.1.15 and was fixed in 0.1.16.

The lesson isn’t “stop using MCP.” It’s that every MCP server you connect to your agents is, in practice, third-party code with permission to act inside your company — and in most deployments, it’s being treated as harmless by default.

What nobody says when they sell “agents connected to everything”

The dominant pitch for AI agents is “the more tools it can reach, the more valuable it is.” Nobody sells the opposite sentence, which is the one that actually protects a company: every MCP server you connect is a new door, and the overwhelming majority of those doors, today, have neither a real lock nor anyone keeping a record of who walked through. The problem isn’t MCP itself — it’s a solid piece of protocol design. The problem is that it’s being rolled out at the pace of a trend, skipping the vendor onboarding, security review and access audit that any other piece of third-party software would have to clear before touching a production system.

What actually governing an MCP Gateway requires

The data above points to five concrete, non-optional controls for any company connecting AI agents to MCP servers:

1. An allow-list of approved MCP servers. No agent connects to a new MCP server without someone reviewing it first — the same way software doesn’t get installed without going through IT. 2. OAuth or an equivalent credential system that expires, never static keys stored in plain text. Fewer than 1 in 10 audited servers do this today. 3. An audit log for every call to every tool: which agent, which server, which action, when — not a generic application log. 4. Network isolation for the MCP server, so a command-injection flaw doesn’t translate into access to the rest of the infrastructure. 5. Periodic review of which servers are still connected, and why — pilots that quietly graduate into production without review are, according to the abuse data itself, one of the most common entry points.

An example of how this fails in practice

An illustrative example, not a real case: a team connects its AI agent to an open-source MCP server that syncs support tickets with a CRM, found in a public repository and tested for an afternoon because “it worked fine.” The server authenticates with a fixed, non-expiring API key, stored in an environment variable inside the agent’s own container. Nobody onboarded it as a vendor or ran a security review, because without “a contract with a company” behind it, none of the usual processes ever triggered. Months later, an update to the server — outside the company’s control — adds a new feature that also reads billing fields in the CRM, not just tickets. Nobody notices, because there’s no log distinguishing what the server touched before and after the change. The failure isn’t in the AI agent itself: it’s in treating a piece of third-party software as if it weren’t one.

At Optima Quantum we don’t connect AI agents to MCP servers without going through this review first: what the server is, who maintains it, what credentials it connects with, and what gets logged for every action. If your company already has AI agents connected to tools through MCP and you couldn’t answer today how many distinct servers that involves or who approved them, that’s exactly what a cybersecurity audit covers. Contact us before you find out through an incident, not after.

Frequently asked questions

What is MCP (Model Context Protocol)? It’s an open standard that lets an AI agent connect to external tools and data sources — email, a CRM, databases, code repositories — in a uniform way, instead of a custom integration for each one. Claude, ChatGPT and Gemini all use it.

Is an MCP Gateway the same as a firewall? Not exactly. A firewall filters network traffic in general; an MCP Gateway specifically controls which agent can talk to which MCP server, with what credentials, which tools it can call, and logs every call — it’s access control and auditing at the protocol level, not the network level.

Why do 73% of MCP servers have vulnerabilities if serious companies back the protocol? Because the protocol itself isn’t the problem — the implementations are. Anyone can publish an MCP server, and most of the ones in active use today were written by an independent developer or a small team, without the review process an enterprise software vendor would go through.

How do you know if AI agents in your company are already using ungoverned MCP servers? By reviewing, server by server, what each deployed agent connects to, what kind of credential it uses, and who approved that connection. If the answer to “who approved this?” isn’t clear, there’s no real governance, incident or not.

Does this only affect companies that build their own AI agents? No. Any company using an AI assistant connected to its internal tools through MCP — even if a third-party vendor built the agent — depends on the security of those servers. The question to ask any vendor is the same: which MCP servers do they use, with what authentication, and who audited them.

Related: that same question needs to be asked of any vendor before they get access to your systems, not only the ones arriving through MCP — we cover that in AI vendor risk monitoring.

What you don’t share is lost forever

Link copied

Get our new articles by email

One email per article. Unsubscribe anytime.