VATAHosted MCP security workspace

The connector is what gets breached.

In late October 2026 a German hosting company was breached through the MCP server it built so customers could attach an AI assistant to their console. The attacker did not break the console, they broke the connector in front of it, and eight days in they had reached 56 customer projects.

An independent assessment of your MCP server records how it handles sign-in, what it exposes without a credential, and what changes when you ship a fix. The report carries a hash a second party can recompute by re-running the same checks on the same server, so your buyer can verify it instead of trusting it.

One server, 1750 dollars. Full assessment with one retest, 3500 dollars.

Independent MCP security assessments

01

Single-server assessment

1750 dollarsBuy the single-server assessment

One MCP server or a small set of endpoints. Written scope and authorization before testing, evidence-backed findings, a written report signed by the named researcher, and one retest of the same endpoint after fixes.

02

Full assessment

3500 dollarsBuy the full assessment

A server or fleet, one written report, one retest, and the written scope agreed by email before testing.

Sample report

A real run, published with the operator's agreement and redacted, from a free scan rather than a paid assessment.

Target:
agentpaystore.com/book/mcp
Server:
agentpay-bookings 1.0.0
Tool:
mcp-authscan 1.1.24
Probe set:
2026.11.16.2
Verdict:
HIGH
Findings:
1
Coverage:
1 of 1 testable
Excluded:
15
Hash:
e08ed15790a437ed93573757e377f153aedd043a430cf1e3de93ddea514063e5

Fifteen probes returned no verdict on this target. Twelve of them share one reason: the server publishes no authorization server, so the OAuth discovery, issuer, token and client-registration checks had nothing to test against. They are recorded as excluded, not passed, and the report says so on its face.

The hash covers graded content only, the probe results, findings, severities and the basis for each severity, and leaves out timestamps and server-varying headers. A second party can re-run the same probe set on the same target and either recompute the same hash or name the fields that differ.

View the sample report

Send the server URL and who authorizes the test. Scope and written authorization are agreed by email before anything is tested. Do not include credentials, tokens or secrets in email.

A report may support a buyer review; acceptance is up to the buyer.

Named researcher signs the report
Written authorization before testing
Scope and limitations stated plainly
Evidence and repeatability

How the report stands up

A report is a claim by someone with an interest in it being believed. The record should make that claim inspectable.

The report carries a hash over the graded content. With written authorization, a second party can re-run the same probe set and compare. The comparison is recorded as MATCH, DIVERGENCE or NOT REPRODUCIBLE.

The method does not claim the tool or report is right; it makes inconsistencies easier to catch. Reproduction depends on the agreed scope, available access, and supported techniques.

Who signs the report

Leroy H. Mason

Independent AI security researcher, credited reporter on published MCP security advisories. The assessment is a scoped technical review, not a certification, guarantee of security, or promise of buyer acceptance.

Free, open-source checks

Run the handshake checks yourself first.

The MIT-licensed mcp-authscan is available on GitHub and PyPI. Its free scanner checks the handshake; paid work adds an authorized tool-call check and a named written report within the agreed scope.

Inspect the tool

Start with the same public checks.

The paid check is performed only with written authorization and where the agreed scope and supported techniques allow it. Neither an authorized tool-call check nor a particular coverage outcome is guaranteed before scope is agreed; no test is represented as already run.

The existing workspace
Run checks. Keep the findings in view.

The MCP AuthScan workspace remains available for teams running authorized scans, reviewing their results, and using the existing scan API.