内容
# mcp-trust-anchor
**Cryptographic context-signing proof-of-concept for MCP trust boundary enforcement.**
Companion repository for the Help Net Security column:
*"Cross-Boundary Context Poisoning: The Hidden Threat in the Model Context Protocol (MCP)"*
by Sunil Gentyala.
---
## The Problem
MCP context payloads carry no provenance metadata. A payload sourced from a public
web scrape is structurally identical to one sourced from a vetted internal knowledge
base. High-trust tools have no mechanism to distinguish them. This creates a direct
path for cross-boundary context poisoning and indirect privilege escalation.
## The Solution
This PoC wraps MCP tool responses in signed JWT envelopes that assert:
- **Trust level** (PUBLIC, INTERNAL, CONFIDENTIAL, RESTRICTED)
- **Originating agent identity**
- **Source URI**
- **SHA-256 payload digest** (tamper detection)
- **Issuance timestamp**
High-trust tools validate this envelope before processing any inbound context.
Payloads that fail signature verification or fall below the tool's minimum trust
threshold are rejected with a structured error.
---
## Repository Structure
```
mcp-trust-anchor/
├── README.md
├── requirements.txt
├── keys/
│ ├── generate_keys.py # ECDSA P-256 key pair generator
│ └── .gitignore # Excludes private keys from commits
├── src/
│ ├── context_signer.py # Signs MCP context payloads as JWT envelopes
│ ├── mcp_validator.py # High-trust tool validation gate
│ ├── trust_levels.py # Trust level enum and policy definitions
│ └── key_registry.py # In-memory public key registry (stub)
├── examples/
│ ├── demo_sign_and_validate.py
│ └── demo_poisoning_rejected.py
└── docs/
├── threat_model.md
└── production_checklist.md
```
---
## Quick Start
```bash
git clone https://github.com/sunilgentyala/mcp-trust-anchor
cd mcp-trust-anchor
pip install -r requirements.txt
# Generate a key pair for a sample agent
python keys/generate_keys.py --agent-id web_scraper_agent
# Run the sign-and-validate demo
python examples/demo_sign_and_validate.py
# Run the cross-boundary poisoning rejection demo
python examples/demo_poisoning_rejected.py
```
---
## Trust Level Policy
| Level | Numeric | Permitted Sources |
|--------------|---------|------------------------------------------------|
| PUBLIC | 1 | External web, unauthenticated APIs, RSS feeds |
| INTERNAL | 2 | Authenticated internal services |
| CONFIDENTIAL | 3 | Restricted internal systems, vetted databases |
| RESTRICTED | 4 | Classified or regulated data stores |
High-trust tools declare a **minimum acceptable level**. Payloads below that
threshold are rejected before any processing occurs.
---
## Signing Algorithm
All signatures use **ECDSA with P-256 (ES256)**. P-256 was selected for:
- Compact signature footprint (64 bytes) relative to RSA-2048
- Native hardware security module (HSM) support across major vendors
- Broad library support (cryptography, python-jose, PyJWT)
**Production note:** This PoC uses file-based PEM keys. Production deployments
should store private keys in an HSM or secrets manager (AWS KMS, Azure Key Vault,
HashiCorp Vault) and implement key rotation with a minimum 90-day cycle.
---
## What This PoC Does Not Cover
This repository is a signal-focused proof of concept. The following are
deliberately omitted to reduce noise:
- Distributed public key registry / PKI infrastructure
- Key rotation and revocation workflows
- Replay attack prevention (nonce registry)
- mTLS channel binding for MCP transport
- Integration with specific MCP server implementations (Claude, LangChain, AutoGen)
The `docs/production_checklist.md` file covers recommended controls for each
of these areas.
---
## Related Projects
- [ContextGuard](https://github.com/sunilgentyala/contextguard) - Zero-trust middleware for MCP security
- [GSH Framework](https://github.com/sunilgentyala/gsh-framework) - Agentic AI threat hunting mapped to MITRE ATLAS and NIST CSF 2.0
---
## Author
**Sunil Gentyala**
Lead Cybersecurity and AI Security Consultant, HCLTech
IEEE Senior Member | Cloud Security Alliance Representative
LinkedIn: [linkedin.com/in/sunil-gentyala](https://linkedin.com/in/sunil-gentyala)
---
## How to Cite
This repository accompanies the Help Net Security column "Cross-Boundary Context Poisoning: The Hidden Threat in the Model Context Protocol (MCP)". To cite the software:
```bibtex
@software{gentyala2026mcptrust,
author = {Gentyala, Sunil},
title = {mcp-trust-anchor},
year = {2026},
version = {0.1.0},
url = {https://github.com/sunilgentyala/mcp-trust-anchor}
}
```
Machine-readable metadata is in [`CITATION.cff`](CITATION.cff); GitHub shows it under "Cite this repository".
---
## License
MIT License. See LICENSE for details.
MCP 市场