Six dollars buys you an MCP server

October 7, 2026 · SPR{K3 Research

OX Security scanned the public MCP marketplaces last week. They counted 15,465 servers across five registries. Six of them sit on expired domain names that anyone can buy right now for four to twelve dollars a year. The agents that already added those servers to their configuration do not know this.

That is the shortest version of "Welcome to the Jungle: What We Found Inside 15,465 Public MCP Servers", published Oct 6 at The Hacker News with the primary research by OX Security. It is the third tier-1 security vendor this quarter to publish the same finding against the MCP tier: there is no bouncer at the door, and the floor underneath the door is not solid either.

What the numbers say

The $6 attack

The expired-domain finding is the one that reads like a vulnerability report. The MCP configuration file an agent reads at startup names tools by hostname. When the agent calls that tool, it resolves the hostname and talks to whatever is at that address.

If the hostname's domain has lapsed, anyone with twelve dollars can buy it. The DNS record now resolves to the new owner. The agent's next tool call reaches a server the attacker controls. The agent believes it is calling the tool its operator configured, because that is the name in its configuration file.

Nothing is exploited. No prompt is injected. No model is poisoned. The configuration was correct the day it was written and became an attacker's by the normal expiration cycle of a domain name. Six domains sit in that state right now, published across five registries, with no vetting path and no takedown process.

This is the shape of the Shai-Hulud npm worm, the HalluSquatting cohort, and the tw-pkgprobe-7731 Twilio-brand squat that ReversingLabs wrote up last month, moved from the dependency registry to the agent-tool registry. Same economics, same price points, different inventory.

The consumer-tunnel finding

Zero point four five percent sounds small until you remember what it means. These are developer laptops behind reverse proxies that terminate at the laptop's local MCP server, routed through ngrok's free tier. There is no production-availability contract. There is no identity-of-operator check at the registry. There is no attestation that the laptop stays reachable, stays updated, or stays owned by the same person week to week.

The MCP marketplace does not distinguish between a production SaaS endpoint and a laptop under someone's desk. From the agent's perspective, both are tools.

The jurisdictional finding

Fifteen point six percent of hostnames resolve outside the US. Nineteen are in China. Eighteen are in Russia. Those are the two counts OX Security called out.

Enterprise buyers who adopted MCP for internal data access read documents like OX Security's and then check their own agent configurations. The registry does not label country of origin. Data-residency policies written for SaaS integrations do not apply at the MCP tier, because the MCP tier does not expose the signals those policies need.

Three vendors, same finding

OX Security is not the first to describe this pattern. Snyk's MCP vulnerability-density scan and Socket's deep-scan of 3,267 agent skills arrived at the same framing earlier this quarter: the marketplace is the attack surface. The Oct 6 piece makes the case with a census.

Three independent tier-1 security vendors have now published variants of "the MCP marketplace does not verify what it indexes" within six months. There is no CVE here — only a market structure in which anyone can publish an MCP server to a registry under any identity, and six of them resolve to domains anyone can buy.

What to do about it

Treat an MCP server in your agent's configuration the way you would treat an unauthenticated HTTP service that writes to your database. Because that is what it is.

The takeaway

Six dollars buys you a hostname. A hostname is what an MCP server is, from the agent's perspective. OX Security's census says six of them are for sale right now, 349 agent skills already reference domains that resolve to scams, and the registries do not know and will not know, because that is not what they do.

The AI-agent industry has standardized on an open marketplace model for tools. The scanner that would notice which tools point somewhere else today than yesterday does not live at the registry. It lives inside the inference loop, where the agent's actual tool calls can be observed against what the operator thought the agent was configured to do.

That is the layer that still holds. The configuration file does not.


SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.