arXiv — MCP-DPT: Defense Placement Taxonomy for Model Context Protocol Security
AI relevance: This research addresses fundamental security gaps in Model Context Protocol infrastructure, which underpins modern AI agent toolchains and autonomous system deployments.
Researchers from Old Dominion University published MCP-DPT (Defense Placement Taxonomy), a comprehensive framework that shifts MCP security analysis from attack-centric to defense-placement oriented. The paper systematically maps 49 MCP-specific attacks across six architectural layers and identifies critical coverage gaps in current security practices.
Key Findings
Six Architectural Defense Layers
- Model Provider/LLM Alignment: Model behavior, refusal logic, tool selection
- MCP Host/Application: Execution state, capability exposure, orchestration logic
- MCP Client/SDK: Protocol parsing, request construction, response interpretation
- MCP Server/Tool Execution: Runtime environment, authentication, isolation
- Transport/Network: Communication channels, encryption, endpoint authentication
- Registry/Marketplace & Supply-chain: Discovery, distribution, versioning, provenance
Critical Security Gaps
- Uneven defense coverage: Current protections cluster around tool-centric defenses
- Transport layer under-defended: Only 2/13 analyzed defenses cover network-level threats
- Supply chain risks: Registry and marketplace layers receive minimal protection
- Host orchestration gaps: Application-level enforcement remains sparse
- Structural rather than incidental: Gaps reflect architectural misalignment, not isolated flaws
Primary vs Secondary Defense Points
The taxonomy introduces a crucial distinction:
- Primary defense layer: Earliest boundary where prevention is feasible
- Secondary defense layer: Fallback containment if primary fails
- Example: Rug pull attacks pass registry evaluation (primary) but require host-level detection (secondary)
Why This Matters
MCP adoption has exploded—Anthropic's protocol hit 97M monthly SDK downloads in January 2026—yet security practices haven't kept pace:
- Multi-party trust boundaries: MCP spans independently operated components
- Pre-execution artifacts: Tool metadata influences decisions before any execution
- Cross-context propagation: Attacks chain across tools, sessions, and servers
- Third-party control: Heterogeneous identity and authorization outside model provider boundaries
Defense Mechanisms Analyzed
The study evaluates 13 academic and industry defenses:
- Static/pre-execution: MCP-Scan, MCPScan.ai, Cisco MCP Scanner
- Behavior-level/runtime: MCIP-Guardian, MCP Defender, MCP-Guard
- Isolation-based/architecture: MCP-Gateway, ToolHive, MCP Guardian
- Decision-level: MindGuard, AIM-Guard-MCP, Prisma AIRS
Recommendations
- Adopt defense-in-depth: Implement protections across multiple layers
- Focus on transport security: Strengthen network-level protections
- Enhance supply chain governance: Implement registry-level controls
- Improve host orchestration: Strengthen application boundary enforcement
- Use capability-based assessment: Evaluate defense coverage across attack classes