← 返回列表

MCP server security testing

单站 Appsecco 入库 2026-11-18T07:15:14

Public research and training

GitHub proof buyers can inspect before they talk to us.

Our MCP security work is public and practitioner-led — including the

Type to search across all pages

Search by Pagefind

MCP Server Pentesting

Your MCP servers connect AI assistants to databases, file systems, and internal APIs. We test whether an attacker can exploit that connection.

Fixed scope. Tool-by-tool evidence. Retest included within 30 days.

157+

GitHub stars on vulnerable-mcp-servers-lab

9

MCP vulnerability categories documented

13

Phase testing methodology

Checklist

Authors of the public MCP pentesting checklist

What's at risk

MCP servers let assistants invoke tools, read resources, and act on connected systems. When those servers are vulnerable, prompt-layer attacks can become database queries, file reads, internal API calls, or workflow actions.

Command injection through tool parameters

Context poisoning via hidden instructions in tool descriptions

Data exfiltration through side channels

Tool shadowing - malicious tool overrides a trusted one

Privilege escalation via tool chaining

Authentication hijacking through OAuth/token flaws

What we test

Each domain maps to a concrete failure mode in MCP deployments: transport trust, unsafe tools, prompt-to-tool abuse, resource exposure, credential handling, and supply chain trust.

1

We test stdio, HTTP, and SSE transport boundaries for tampering, replay, exposure, and unsafe trust assumptions.

2

We exercise every tool parameter for command injection, traversal, SSRF, unsafe parsing, and missing validation.

3

We trace whether malicious content can change model behavior, expose secrets, or leak data through tool outputs.

4

We verify which files, databases, APIs, and tenant resources the server can reach and whether boundaries hold.

5

We review scopes, token storage, refresh flows, logs, and multi-user isolation for credential abuse paths.

6

We check package provenance, dependency exposure, tool registration integrity, and malicious-server assumptions.

Our credibility

Our assessment approach is grounded in public research, intentionally vulnerable labs, and hands-on tooling for real MCP server behavior. Every test plan is protocol-specific — not a standard application checklist with MCP wording added later.

Tool poisoning Rug pull attacks Tool shadowing Command injection Prompt injection Context poisoning Data exfiltration OAuth abuse Supply chain compromise

See the deliverable

Review the protocol-specific depth and reporting quality before you commit to an MCP assessment.

Report format

MCP engagements follow the same reporting discipline as our product security work, with tool-by-tool matrices, prompt-to-tool traces, and connected-resource notes added where the protocol creates extra risk.

How it works

We start with the actual servers and tools you run, then test the places where AI interpretation meets system access.

Step 1

Enumerate servers, tools, resources, transport, auth, and runtime boundaries.

What happens

We build an inventory of MCP servers, exposed tools, resource permissions, trust boundaries, transport modes, and authentication flows.

What you do

Share the server list, access paths, architecture notes, and rules of engagement.

What we do

Confirm the test matrix and mark the highest-risk tool and data paths before active testing starts.

What comes next

A clear assessment map anchors every finding to the server, tool, and resource it affects.

Step 2

Run injection, traversal, SSRF, and unsafe parsing checks on each parameter.

Each tool is exercised with adversarial inputs, malformed requests, boundary bypass attempts, and chained tool-call scenarios.

Provide safe test data or staging access where destructive behavior must be avoided.

Record reproducible evidence and separate exploitable issues from defensive noise.

You receive a tool-by-tool matrix that makes remediation ownership clear.

Step 3

Probe prompt injection at every stage of the prompt, resource, tool, and response pipeline.

We test whether hidden instructions, tool descriptions, resource content, and retrieved data can alter behavior or leak information.

Identify sensitive data classes, tenant boundaries, and content sources in scope.

Trace attack paths through model context, tool outputs, side channels, and downstream systems.

Findings show how data moves, where trust is misplaced, and how to reduce exposure.

Step 4

Assess secret storage, token handling, OAuth flows, scopes, and tenant isolation.

We inspect how credentials are stored, passed, logged, refreshed, scoped, and isolated across users and servers.

Share the intended permission model and any constraints for tokens or connected services.

Look for scope creep, token leakage, replay paths, weak OAuth assumptions, and auth confusion.

Credential findings include least-privilege recommendations and validation steps.

Step 5

Audit dependencies, package provenance, and tool registration integrity.

We review installed MCP packages, dependency vulnerabilities, malicious-server assumptions, and whether registered tools can be trusted.

Provide package manifests, deployment details, and approved source locations.

Check provenance, dependency risk, update posture, and integrity controls around server registration.

You receive a defensible inventory and prioritized fixes for trust and dependency gaps.

3-5 days for a single server

5-10 days for multi-server ecosystems

Retest included within 30 days

MCP testing matters when AI assistants can reach systems that were never designed to be prompt-facing.

You are shipping integrations to customers and need confidence that exposed tools cannot be abused.

You receive:

Tool-by-tool assessment matrix with reproducible findings

You are connecting assistants to internal tools, files, databases, or workflows used by your team.

Access boundary review and configuration recommendations

You are embedding AI capabilities in a product and need evidence for customers, security, or leadership.

Integration security report with attack-path narratives

Pricing

Scope depends on the number of servers, exposed tools, connected resources, tenant model, and auth complexity.

Scope

Price

Duration

Single MCP server (\< 10 tools)

$3,500-$5,000

3-5 days

Multi-server (2-5 servers)

$7,500-$12,500

5-7 days

Enterprise (5+ servers, multi-tenant)

From $15,000

7-10 days

Add-on to existing pentest

$2,000-$3,500

1-2 days

Fixed price. No hourly. Quote in 48 hours. Retest included within 30 days.

What you get

The output is built for remediation, review, and proof. You get the attack path, the affected tool or resource, and the specific change needed to close the issue.

Executive summary with prioritized findings

Technical report with attack-path narratives

Tool-by-tool assessment matrix

Remediation guidance specific to your MCP framework

Supply chain audit results

Verification letter / attestation on request

Walkthrough call included

One free retest within 30 days

We test the server transport, every exposed tool, prompt-to-tool data flow, resource boundaries, OAuth and token handling, and the supply-chain assumptions around installed MCP packages and registered tools.

Both. We assess internally built MCP servers as well as framework-based deployments, provided they implement the protocol and can be exercised in a controlled environment.

Yes. If your product includes both model-facing application logic and MCP servers, we can scope them together so the report covers the full prompt, tool, auth, and data path.

We usually start with staging access, architecture notes, and enough credentials to exercise the in-scope tools. If production validation is necessary, the exact boundaries and safe methods are agreed before testing starts.

You receive a report with prioritized findings, tool-by-tool coverage, attack-path narratives, remediation guidance tied to your MCP stack, and a retest window so fixes can be verified.

Explore the MCP security surface

Continue from the concept into testing scope, implementation risks, tools, and adjacent AI security topics.

Service

Product security testing for AI apps, agent workflows, MCP tools, prompts, and connected data sources.

Guide

A practical guide for evaluating MCP assessment scope, vendor depth, reporting quality, and protocol-specific coverage.

Clarify when the risk lives in the agent workflow, the MCP server boundary, or both.

Glossary

How malicious instructions enter prompts through users, documents, retrieved content, and tool output.

Security controls for agents that use tools, memory, approvals, and connected workflows.

Risks and controls for LLM applications, RAG systems, embeddings, and model-connected workflows.

Tool

Open-source checklist for reviewing MCP server security, tool safety, auth boundaries, and data exposure paths.

Appsecco testing client and proxy for exercising MCP servers during security reviews.

Intentionally vulnerable MCP servers for learning attack paths and validating defensive controls.

Safe next step

before someone else does

Share what your MCP servers can reach and how they are used. We will outline a scoped assessment, answer questions, and give you a fixed quote before any work begins.

or download the MCP pentesting checklist first

No sales pressure

Fixed pricing

You decide pace

原文链接:https://appsecco.com/mcp-pentesting