cd /news/ai-safety/from-9-to-63-letting-a-benchmark-bui… · home topics ai-safety article
[ARTICLE · art-62779] src=dev.to ↗ pub= topic=ai-safety verified=true sentiment=↑ positive

From 9% to 63%: Letting a Benchmark Build a Better MCP Security Proxy

A developer built mcp-defense-bench, an open benchmark for evaluating security proxies for the Model Context Protocol (MCP), which allows AI assistants to interact with external tools. Starting from a 9% coverage baseline, the developer iteratively improved the mcp-bastion proxy to 63% coverage by addressing 24 attack vectors, including novel attacks like ShareLock that no other tool detected. The work demonstrates a measurement-driven approach to hardening AI-agent security infrastructure.

read6 min views1 publishedJul 16, 2026

A year ago, "AI security" mostly meant "don't let the chatbot say something dumb." Today, AI assistants

act: they read your files, query your databases, hit APIs, and browse the web. They do it through a

fast-spreading open standard called the Model Context Protocol (MCP) — think of it as USB for AI

agents, one standard plug between an assistant and the tools it uses.

That plug is powerful, and it's a brand-new place for things to go wrong. A malicious or compromised

tool can hide instructions in its own description, quietly swap its behavior after you've approved it,

or trick the agent into shipping your credentials to an attacker. The 2025–2026 research literature now

documents dozens of these attacks.

So the obvious question: when you put a security tool in front of your MCP servers, how much of that danger does it actually stop? Nobody had a good number. I set out to get one — and then something more

mcp-bastion

This is the story of that loop, and what it teaches about where AI-agent security actually stands.

I built mcp-defense-bench, an open, vendor-neutral The rubric behind it is a threat–control crosswalk: 24 MCP attack vectors, each mapped to an

architectural layer, a STRIDE class, and the governance frameworks organizations actually use — the

NIST AI Risk Management Framework, the NSA's MCP Security guidance, and the OWASP Top 10s for LLM and

Agentic applications. (If you want the "map of the whole landscape" version of this story, that's a

a separate write-up; here I'm focused on the improvement loop.)

Each attack scores one of three levels: none (missed), detect (warns you), or enforce (blocks it). Weighted coverage counts an enforce as 1.0, a detect as 0.5.

Here's the part that surprised me. A benchmark is usually a scoreboard you check at the end. This one

became a to-do list.

The proxy started at 9% — it scanned tool descriptions for poisoning and nothing else. Then the

benchmark pointed at the next gap, I closed it, and re-ran. Every step below was re-verified at zero false positives:

Step What the benchmark exposed → what I added Coverage
baseline description scanning only 9%
1 result-borne injection → scan tool results, not just definitions
18%
2 parameter smuggling → validate arguments against the tool's schema 23%
3 DNS-rebinding on local servers → an origin check that blocks
30%
4 cross-tool exfiltration → sensitive-argument + least-privilege scanning 34%
5 a brand-new split-payload attack → cross-tool correlation 38%
6 four vectors no tool covered → command-injection, config-drift, server-identity, cross-server taint
48%
7 depth, not breadth → evasion normalization + inline secret redaction + block-by-default 63%

Two moments in that table are worth slowing down on, because they're where "measure a defense" turned

into "discover something."

Halfway through, two attacks were published (June 2026) that my frozen rubric didn't have yet:

mid-session tool injection (a tool that mutates its behavior after you've approved it) and

ShareLock (a stealthy attack that splits a malicious payload across several tools, so no single

tool description looks suspicious).

I verified both against their source papers, encoded each as a matched attack/benign pair, and re-ran

every tool. The result was stark: ShareLock was covered by nothing. Every scanner looked at one tool

at a time and saw nothing wrong.

But a proxy sees the whole tool set at once. So I added cross-tool correlation — it scans the

combined descriptions and flags the coordinated "share/checksum"-style staging metadata the attack uses.

mcp-bastion

became the first and (on the benchmark) only measured tool to catch it. A benchmark that

only tested a fixed set of attacks would have missed this entirely. The value was in re-measuring as the threat surface moved.

Later, I added cross-server data-flow tracking. Picture an agent reading a secret from your vault

server and then, two calls later, posting it to a webhook on a different server. No single tool did

anything wrong; the danger is in the flow across a trust boundary. A per-server scanner is blind to

it. A proxy that fronts all your servers isn't — it can watch a credential-shaped token read from server

A reappear in an argument headed to server B, and stop it. That's not a rule you can bolt onto one

tool; it's a capability you only get from the vantage point.

A tool that detects an attack and prints a warning is a linter. A tool that blocks it is a firewall.

The last leg of the journey (48% → 63%) was mostly about turning warnings into blocks — safely.

Through all of it, the false-positive count stayed at zero across 35 test cases. That's the whole

point of the matched benign controls: you can push toward "block by default" because you've proven the

checks don't fire on innocent traffic.

Here's the result I care about most, and it's not the 63%. It's the 5 of 24 vectors that no measured tool covers — and that no content-scanning proxy ever will:

A proxy is necessary but not sufficient. The data makes the case for defense-in-depth across distinct mechanism classes better than any argument could: 19 of 24 vectors are caught by

The loop is no longer manual. mcp-bastion

's continuous-integration pipeline runs the full benchmark

on every change and fails the build if coverage drops below a floor or any false positive appears.

The measurement isn't a one-time paper result; it's a standing regression test on the defense. When the

next MCP attack drops, the process for handling it is already wired: add the vector, re-measure, close

the gap, and the gate makes sure it stays closed.

Everything is open:

mcp-defense-bench mcp-bastion

npx -y mcp-bastion ) and in the official MCP registry.One disclosure, up front: I build one of the tools the benchmark measures. That's exactly why the

whole thing is open, vendor-neutral, driven through each tool's real code, and scored alongside

competitors and a do-nothing baseline — so you don't have to take my word for any of it. Re-run it

yourself.

The number that matters isn't 63%. It's that an honest measurement, applied relentlessly, turned a

9%-coverage scanner into the broadest MCP defense measured — and drew a bright line around the 37% that

needs a fundamentally different kind of tool. That line is the useful part.

If you work on MCP, agent security, or AI governance, I'd genuinely value critique of the rubric and the methodology — open an issue, or reach me at agowthaman90@gmail.com.

── more in #ai-safety 4 stories · sorted by recency
── more on @mcp-defense-bench 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/from-9-to-63-letting…] indexed:0 read:6min 2026-07-16 ·