ENFORCEAUTH 2027 — NO. 05 OF 18 — MCP
By Mark Rogge, CEO of EnforceAuth | 6 min read | for security architects putting agents in front of enterprise systems
In February 2026, Trend Micro counted 492 Model Context Protocol servers reachable on the open internet with no authentication at all. By the third quarter the count was roughly 200,000. In the same period Elastic tested a sample of MCP servers and found command-injection flaws in 43% of them, and more than 30 MCP-related CVEs were published in a single 60-day window.
The response from the industry is already visible: put a gateway in front of it. By the end of 2027 I expect an MCP gateway to be a line item in every regulated enterprise's agent architecture. I also expect the 2027 incident reports to show that the gateway was in place and the damage happened anyway. This post is about why.
What a gateway is for
A gateway answers a connectivity question. Which MCP servers may this agent reach? Is the server authenticated? Has the tool manifest been scanned? Are the parameters well-formed? Those are good questions and the NSA and CISA MCP Security advisory of June 2, 2026 is a useful list of what to check for: tool poisoning, rug-pull redefinition of a tool after approval, parameter injection.
The 2026 record shows each one live. DuneSlide (CVE-2026-50548 and CVE-2026-50549, CVSS 9.8, reported by Cato in July) escaped the Cursor sandbox through injected text. Deadbugz, reported in August, delivered a malicious MCP server through GitHub pull requests. In April, the Azure DevOps MCP authentication bypass (CVE-2026-32211, CVSS 9.1) let callers skip the check entirely. In September the MCP Python SDK leaked OAuth secrets. And Tenet's "Agentjacking" research in June found 2,388 organizations with injectable Sentry DSNs and an 85% execution rate through the Sentry MCP server; Sentry declined a full fix.
A gateway helps with most of those. It does nothing about the following case.
The approved tool does the unapproved thing
Picture a finance agent with access to an approved, authenticated, scanned MCP server that exposes a create_payment tool. The gateway has allowlisted the server. The manifest is clean. The parameters are well-formed. The agent, following a poisoned instruction or a misread ticket, calls create_payment for a vendor it has never paid, for an amount above the agent's delegated authority, at 02:14 on a Sunday.
Every layer in the gateway says yes. The server is approved. The tool is approved. The call is syntactically valid. The one question nobody asked is whether this identity may perform this action on this resource, now, under whose authority.
Tool connectivity answers what exists. Authorization answers this action, now, under whose authority. A gateway is the first. It is routinely sold as the second.
The counter-argument I hear from gateway vendors is that per-tool allowlists with parameter rules handle this. They handle static cases. They do not know that the agent's authority was delegated by a user whose own limit is $5,000, that the vendor was added to the master file four minutes ago by the same agent's sub-agent, or that the ALLOW issued at the start of the session should be revisited because a privileged role changed in between. That is state, relationships and time. A request filter does not hold them.
The 2027 forecast, in three parts
Gateways become table stakes. By the end of 2027, allowlisting, authentication and manifest scanning for MCP will be a baseline expectation in bank and insurer architecture reviews, the way an API gateway is today. OWASP's 2026 Top 10 for LLM Applications already cites MCP under LLM01 prompt injection and separately lists Excessive Agency.
Exposed servers keep growing. The 492-to-200,000 curve in nine months is not slowing. Most of the growth is developer convenience, not production design, which is why HackerOne reports a 540% surge in prompt-injection submissions.
The incidents move up a layer. The disclosures that matter in 2027 will describe an allowlisted tool, called by an authenticated agent, through a scanned server, that executed a transaction outside the agent's authority. Surveys already put 47% to 53% of organizations as having had an agent exceed its permissions. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, and I expect security review to be the stated reason in a large share of them.
Where EnforceAuth sits, and what ships
We build the layer between the tool call and the resource. Concretely, today:
- Verdict is our authorization MCP server. An agent's tool call is evaluated by an OPA and Rego policy decision point against the identity, the resource, the action, the delegation chain and the current state. ALLOW or DENY, with reasons, before the resource is touched.
- OpenID AuthZEN is the API we expose, so a gateway, an IdP or an application can ask the same question in a standard way. We are complementary to the gateway, not a replacement for it.
- EvidenceIQ, part of the EnforceAuth platform, holds the context the request filter cannot: the relationships between the user, the agent, the sub-agent, the vendor record and the approval, and it re-evaluates after the first ALLOW when the graph changes. Every decision is recorded, tied to the policy version, replayable, retained one to four years, exportable to Splunk, Datadog or S3.
- Deployment is multi-tenant SaaS, single-tenant SaaS, BYOC through Terraform or Pulumi, or air-gapped on-prem, because a bank's
create_paymentpolicy is not leaving the bank.
Design targets for the PDP are p50 under 2ms and p99 under 5ms; those are goals, not measurements. Verdict, our agentic firewall, is coming and is not what this post describes.
Put the gateway in. Then put the decision behind it.
EnforceAuth is the runtime authorization layer for AI agents, built on OPA and Rego. See enforceauth.com and read the EvidenceIQ announcement.
About EnforceAuth
EnforceAuth is the OPA/Rego authorization control plane. The customer's OPA decides what people, services, and AI agents may do at action time. Verdict covers high-impact agent actions. EvidenceIQ keeps the record behind every decision.

