VECTOR

MCP Protocol Exploitation

Every tool call is an attack surface. VECTOR finds the gaps. Tool description poisoning, parameter injection, SSRF, data exfiltration, server impersonation, authentication bypass, and registry poisoning — weaponised for authorised red team engagements.

7
Subsystems
75
Attack Vectors
172
Tests
View Documentation GitHub

Nobody Tests the Protocol

The Model Context Protocol is the emerging standard for AI tool integration. Every major AI framework is adopting it. Nobody is testing whether the implementations are secure. MCP servers are deployed with no authentication, no input validation, no output sanitisation, and no integrity verification. Every tool call is an attack surface — and nobody is looking.

Blind Trust in Tool Descriptions

AI agents read tool descriptions and follow them without question. A poisoned description tells the agent to exfiltrate data, read sensitive files, or override its own safety instructions. The agent complies because the description says to. No verification. No integrity check. Just trust.

Unsanitised Parameters

MCP tool parameters flow from AI agents to system calls with zero sanitisation. Shell injection, path traversal, SQL injection, template injection, argument injection — the entire classic web application attack surface is exposed through a protocol that was never designed for adversarial input.

No Authentication Standard

MCP has no mandatory authentication. Servers accept any connection. Tokens are optional. JWT validation is rare. The protocol specification focuses on capability, not security. The result: 38% of MCP servers have zero authentication.

SSRF by Design

MCP servers make HTTP requests on behalf of AI agents. Cloud metadata endpoints, internal services, Redis, Elasticsearch, Kubernetes APIs, Docker sockets — all reachable through crafted tool parameters. The MCP server becomes an unwitting proxy into your internal network.

Invisible Supply Chain

MCP tool registries have no integrity verification. Typosquatting, version confusion, publisher impersonation, dependency confusion — every supply chain attack that plagued npm and PyPI now applies to MCP tool registries. Nobody is checking.

No Existing Tooling

Burp Suite does not understand MCP. OWASP ZAP does not understand MCP. Metasploit does not understand MCP. There is no security testing tool purpose-built for the Model Context Protocol. VECTOR is the first.

Every Tool Call Is an Attack Surface

VECTOR targets the Model Context Protocol — the emerging standard for AI tool integration. Every MCP server, every tool description, every parameter schema, every authentication flow — all exploitable. VECTOR finds the gaps between what MCP promises and what it actually secures.

01

POISONER

14 PAYLOADS • 6 CATEGORIES

Inject malicious instructions into MCP tool descriptions. Manipulate AI behaviour through crafted tool metadata. SHA-256 fingerprinting for tamper detection.

  • Instruction injection (trailing, body, conditional)
  • Description mutation (scope expansion, permission escalation)
  • Parameter redefinition (hidden params, type confusion)
  • Shadow tool registration and name collision
  • Schema manipulation (required field removal, validation bypass)
  • Zero-width Unicode obfuscation detection
02

INJECTOR

20 PAYLOADS • 6 CATEGORIES

Inject payloads through MCP tool parameters. Exploit insufficient input validation across every classic injection class.

  • Shell injection (command substitution, backtick, pipe, newline)
  • Path traversal (basic, URL-encoded, double-encoded, null byte, Windows)
  • SQL injection (UNION, blind boolean, stacked query, time-based)
  • Argument injection (git, curl flag injection)
  • Template injection (Jinja2 SSTI, expression language)
  • Environment variable expansion
03

SSRF

15 TARGETS • 3 CATEGORIES

Force MCP servers to make unintended requests. Cloud metadata, internal services, DNS rebinding — all probed systematically.

  • AWS IMDSv1 metadata and IAM credential theft
  • GCP metadata and service account tokens
  • Azure IMDS and DigitalOcean metadata
  • Internal Redis, Elasticsearch, Kubernetes API, Docker socket
  • DNS rebinding (0.0.0.0, IPv6, decimal IP, hex IP)
  • Dangerous protocol detection (file://, gopher://, dict://)
04

EXFIL

6 VECTORS • 6 CHANNELS

Extract sensitive data through MCP tool responses. Tests every exfiltration channel the protocol exposes.

  • Sampling channel data leak (createMessage abuse)
  • Tool output side-channel (base64-encoded secrets in params)
  • Error message data leak (traceback, stack trace, internal paths)
  • Timing side-channel (bit-by-bit encoding via response latency)
  • DNS exfiltration via tool URL parameters
  • Parameter reflection exfiltration
05

IMPERSONATOR

5 VECTORS • SHA-256 FINGERPRINTING

Impersonate legitimate MCP servers. Server profile fingerprinting, tool list diffing, and trust chain validation.

  • Tool list manipulation (injected tools detection)
  • Response manipulation (MITM tool call interception)
  • Capability downgrade (security feature stripping)
  • Certificate impersonation (hostname verification bypass)
  • Protocol version downgrade
06

AUTH

10 VECTORS • 6 CATEGORIES

Bypass MCP authentication mechanisms. Tests every weakness in the authentication layer — from zero auth to JWT algorithm confusion.

  • No authentication header (zero-auth detection)
  • Empty and null bearer token bypass
  • JWT none algorithm attack
  • Expired token acceptance
  • CORS misconfiguration (origin reflection)
  • Admin endpoint exposure, replay attack, credential leak
07

REGISTRY

5 VECTORS • LEVENSHTEIN DISTANCE

Poison MCP tool registries. Supply chain attacks on tool packages with computational typosquat detection.

  • Typosquatting detection (edit distance + 9 transform patterns)
  • Version confusion (malicious version number escalation)
  • Publisher impersonation (confusable name detection)
  • Dependency confusion (internal vs. public package conflict)
  • Post-install integrity bypass (time-of-check to time-of-use)
7
Subsystems
75
Attack Vectors
172
Tests
14
Poison Payloads
20
Injection Payloads
50,914
Ecosystem Tests

One Command Per Subsystem. Or One For All Seven.

Full scan — all 7 subsystems against a target MCP server:

$ vector scan --target http://mcp-server:3000 --output reports/ --verbose
[VECTOR] Red Specter VECTOR v1.0.0
[SCAN]   Target: http://mcp-server:3000
[SCAN]   Mode: Full scan — all 7 subsystems

  -> poisoner       14 payloads | 6 categories
  -> injector       20 payloads | shell, path, SQL, template, argument
  -> ssrf           15 targets  | AWS, GCP, Azure, internal, DNS rebind
  -> exfil          6 vectors   | sampling, output, error, timing, DNS
  -> impersonator   5 vectors   | tool list, response, capability, cert
  -> auth           10 vectors  | no-auth, JWT none, CORS, replay
  -> registry       5 vectors   | typosquat, version, publisher, dep

SCAN COMPLETE | Risk Grade: D- | 23 findings | Report signed
  JSON: reports/RSV-SCAN-A3F8B1C02D4E_VECTOR_2026-03-26.json

Or target individual subsystems:

$ vector poisoner --target http://mcp-server:3000 --verbose
$ vector injector --target http://mcp-server:3000
$ vector ssrf --target http://mcp-server:3000
$ vector exfil --target http://mcp-server:3000
$ vector impersonator --target http://mcp-server:3000
$ vector auth --target http://mcp-server:3000
$ vector registry --target http://mcp-registry.internal
$ vector report --input reports/ --format json

Poisoning simulation — test detection against all 14 payloads:

$ vector poisoner --target http://mcp-server:3000 --verbose
[VECTOR] Red Specter VECTOR v1.0.0
[POISONER] Target: http://mcp-server:3000
[POISONER] Payloads: 14 poisoning vectors

             Poisoning Simulation Results
ID         Name                          Category            Severity   Status
VPOI-001   Description trailing instruction  instruction_inject   CRITICAL   VULNERABLE
VPOI-002   Parameter description injection   instruction_inject   CRITICAL   VULNERABLE
VPOI-003   System prompt override           instruction_inject   CRITICAL   VULNERABLE
VPOI-004   Unicode obfuscated instruction    instruction_inject   HIGH       VULNERABLE
VPOI-010   Scope expansion mutation         description_mutate   HIGH       VULNERABLE
VPOI-030   Shadow tool registration         shadow_tool          CRITICAL   RESISTANT

Completed in 0.42s

Every Finding Cryptographically Signed

VECTOR produces engagement-grade evidence. Every finding is cryptographically signed with Ed25519. Every evidence item is chained with SHA-256. Every report carries an RFC 3161 timestamp. Tamper with one finding and the entire chain breaks. Designed for court, not just dashboards.

Ed25519

Digital Signatures

  • Ed25519 private key signing
  • PEM key storage (PKCS8 format)
  • 0600 file permissions on private keys
  • Canonical JSON serialisation
  • Signature verification API
  • One operator. Founder's machine only.
SHA-256

Evidence Chain

  • SHA-256 hash chain (blockchain-style)
  • Each entry links to previous hash
  • Genesis block: 64 zero bytes
  • Tamper-evident by design
  • Full chain verification API
  • Append-only evidence collection
RFC 3161

Report Integrity

  • ISO 8601 UTC timestamps on every finding
  • RFC 3161 timestamp support
  • Unique scan IDs (RSV-SCAN-xxx)
  • Unique finding IDs (RSV-xxx)
  • JSON + HTML report formats
  • Machine-ingestible structured output

9 Sensitive Data Patterns. Automatic Detection.

VECTOR's exfil engine scans every MCP tool response for 9 categories of sensitive data leakage. Regex pattern matching detects API keys, bearer tokens, AWS credentials, private keys, JWT tokens, passwords, connection strings, internal paths, and private IP addresses — automatically, on every scan.

Credential Patterns
  • > API Key — api_key/apikey followed by 20+ char value
  • > Bearer Token — Authorization header extraction
  • > AWS Key — AKIA prefix detection (IAM access keys)
  • > Private Key — PEM header detection (RSA, EC, DSA)
  • > JWT Token — eyJ prefix with three-part dot structure
Infrastructure Patterns
  • > Password — password= followed by 8+ char value
  • > Connection String — mysql://, postgres://, mongodb://, redis://
  • > Internal Path — /etc/, /var/, /home/, /root/, /opt/
  • > Private IP — 10.x, 172.16-31.x, 192.168.x ranges
  • > Context data cross-reference against agent memory

10 Dangerous Ports. 8 Internal URL Patterns. 3 Dangerous Protocols.

VECTOR's SSRF engine does not just probe — it understands network topology. Built-in detection for dangerous service ports, internal IP ranges, metadata endpoints, and protocol scheme abuse. Every URL parameter is analysed before any probe is sent.

6379
Redis
9200
Elastic
2375
Docker
5432
Postgres
3306
MySQL
27017
MongoDB
6443
K8s API
11211
Memcached
9300
ES Transport
2376
Docker TLS

Wilson Score Confidence Intervals. Not Vibes.

VECTOR scores every MCP server with weighted severity analysis and statistical confidence intervals. Five severity levels. A+ to F grading with 13 grade thresholds. Wilson score intervals on vulnerability rates. Scipy-backed statistical rigour — not percentage guesses.

10.0
CRITICAL Weight
7.0
HIGH Weight
4.0
MEDIUM Weight
2.0
LOW Weight
0.5
INFO Weight
A+ to A-
Score 0 – 15
B+ to B-
Score 15 – 45
C+ to C-
Score 45 – 75
D+ to F
Score 75 – 100

LEVIATHAN Assesses. VECTOR Exploits.

Red Specter LEVIATHAN performs MCP server security assessments — configuration audits, compliance checks, risk scoring. VECTOR is different. VECTOR is the red team tool that actually exploits MCP servers. It injects, poisons, impersonates, exfiltrates, and bypasses. Assessment tells you what could go wrong. VECTOR proves it.

LEVIATHAN
  • Configuration audit
  • Compliance scoring
  • Risk assessment
  • Passive analysis
  • What could go wrong
VECTOR
  • Active exploitation
  • 75 attack vectors across 7 subsystems
  • Proof-of-exploitation evidence
  • Ed25519 signed forensic reports
  • What actually breaks

VECTOR Exploits. The Pipeline Finishes.

VECTOR is one stage in the NIGHTFALL AI attack pipeline. VECTOR compromises the MCP protocol layer. LAZARUS persists in memory. SERPENT poisons reasoning chains. JANUS bypasses guardrails. Together, they own the entire AI tool integration stack.

Stage 1 — Protocol
VECTOR
Exploit MCP servers, tool descriptions, parameters, authentication, registries
Stage 2 — Persistence
LAZARUS
Persist in AI memory. Survive context resets. Maintain footholds across sessions.
Stage 3 — Reasoning
SERPENT
Poison chain-of-thought reasoning. Corrupt logic without triggering safety.
Stage 4 — Guardrails
JANUS
Bypass safety guardrails. Dual-face attacks. Make the model contradict itself.
Chain: VECTOR + LAZARUS

Compromise MCP tools, then persist the backdoor in AI memory. The agent keeps using the poisoned tool across sessions.

Chain: VECTOR + SERPENT

Poison tool descriptions to inject reasoning corruption. The agent's chain-of-thought follows attacker-controlled logic.

Chain: VECTOR + JANUS

Use MCP tool responses to feed guardrail bypass payloads. The tool output becomes the jailbreak delivery mechanism.

Enterprise Integration
Enterprise SIEM Integration — Native

Export every finding directly to your SIEM. Native format translation. Ed25519 signatures and RFC 3161 timestamps preserved across every export. Every finding includes tool name, subsystem, severity, risk score, grade, payload, response, remediation, and full metadata.

Splunk
HEC • CIM Compliant
Sentinel
CEF • Log Analytics API
QRadar
LEEF 2.0 • Syslog
vector scan --target http://mcp-server:3000 --export-siem splunk
Pure Engineering
Purpose-Built MCP Exploitation Engine

Every payload, every detection pattern, every scoring algorithm, every evidence chain — written from scratch in pure Python. Pydantic configuration. httpx for async MCP client communication. Typer CLI. Rich terminal output. Zero subprocess calls. Zero external tool dependencies.

75
Attack Vectors
9
CLI Commands
3
MCP Transports
0
External Dependencies

UNLEASHED Gate

Standard mode detects. UNLEASHED exploits. Ed25519 crypto. Dual-gate safety. One operator. Every UNLEASHED finding is tagged with [UNLEASHED] prefix, dry-run flag, and full audit trail via the override context logger.

Detection

Maps MCP attack surfaces. Identifies vulnerable tool descriptions, parameter schemas, authentication gaps, and registry weaknesses. No exploitation. Reports only.

Dry Run

Plans full MCP exploitation campaigns. Shows exactly what would work — poisoned descriptions, injectable parameters, SSRF routes, exfiltration paths. Ed25519 required. No execution.

Live Execution

Cryptographic override. Private key controlled. One operator. Founder's machine only. Full exploitation with signed evidence chain and SIEM export.

THIS TOOL IS FOR AUTHORISED SECURITY TESTING ONLY. EVERY EXECUTION IS SIGNED AND LOGGED.

Every MCP Transport. Every Endpoint.

VECTOR supports all three MCP transport mechanisms. stdio for local servers. Server-Sent Events for streaming connections. Streamable HTTP for the latest MCP specification. Configurable authentication headers, custom endpoints, concurrency control, and request delays for stealth scanning.

STDIO

Local MCP servers. Process stdin/stdout communication. Default transport for most MCP implementations.

SSE

Server-Sent Events. Streaming HTTP connections. Long-running tool calls with real-time progress updates.

STREAMABLE HTTP

Latest MCP spec. Full HTTP request/response with streaming support. Configurable endpoints: /tools, /execute, /sampling.

Authorised Use Only

Red Specter VECTOR is intended for authorised security testing only. Unauthorised use against systems you do not own or have explicit permission to test may violate the Computer Misuse Act 1990 (UK), Computer Fraud and Abuse Act (US), and equivalent legislation in other jurisdictions. Always obtain written authorisation before conducting any security assessments. Every VECTOR execution is Ed25519 signed and logged. Apache License 2.0.

Available On

Security Distros & Package Managers

Kali Linux
.deb package
Parrot OS
.deb package
BlackArch
PKGBUILD
REMnux
.deb package
Tsurugi
.deb package
PyPI
pip install
macOS
pip install
Windows
pip install
Docker
docker pull

Every Tool Call Is an Attack Surface. VECTOR Finds the Gaps.

7 subsystems. 75 attack vectors. 172 tests. MCP protocol exploitation. The tool that proves your AI tool integrations are not safe.