Research index

MCP Security Research

Technical threat analysis for MCP security, prompt injection, adversarial RAG, deterministic proxies, and runtime controls for AI agents.

Research posture
Boundary
Before execution
Method
Threat model + runtime evidence
Output
Controls teams can inspect
Runtime enforcement
MCP boundary controls

Policy before execution, explicit verdicts, and safe-zone decisions.

Agent threat models
Prompt, RAG, tool abuse

How language becomes capability, and where attackers insert leverage.

Evidence architecture
Audit-ready records

Decision trails security teams can inspect, reproduce, and govern.

Latest release analysis

Layered MCP Runtime Enforcement

The newest McpVanguard research-backed release explains why semantic scoring is useful as a signal, but dangerous as the security boundary.

Read release analysis
Vulnerability Analysis

MCP Prompt Injection

How adversarial instructions embedded in prompts, retrieved content, or tool outputs can redirect autonomous agent intent — and what deterministic controls actually prevent it.

View Analysis
Threat Model

Adversarial RAG & Context Poisoning

When retrieval-augmented generation becomes a delivery mechanism for malicious context. Exploring poisoning vectors, persistence, and proxy-level countermeasures.

View Analysis
Implementation

Deterministic MCP Proxy

How McpVanguard enforces policy at the MCP boundary without trusting the model. Rule-based interception, schema validation, and real-time audit logging.

View Analysis
Research Collaboration

Work with us

Exploring formal verification, protocol-level trust, or execution governance? We collaborate with security teams and researchers on real deployments.

Get in touch