security MCP Server
The Security Kernel for AI Agents — MCP/A2A gateway with policy enforcement, taint tracking, sandboxed execution, deterministic envelopes, and Sigstore audit. OWASP ASI 2026 compliant.
Discovered via github-topic:model-context-protocol and last synced 3mo ago.
1. Install the package
pip install mcpkernel
key
How It Works
Description
Audit an OpenClaw/ClawHub SKILL.md for dangerous patterns
Find all MCP configurations on the system
Check text for leaked secrets, PII, and API keys
Run health diagnostics on the MCPKernel installation
What MCPKernel Does
What It Does
Scan a tool's description for poisoning, shadowing, and prompt injection
wasm
Validate a YAML policy file for syntax and logic errors
off policy: default_action: deny # deny-by-default for production policy_paths: - policies/owasp_asi_2026_strict.yaml observability: log_level: INFO metrics_enabled: true otlp_endpoint: "" # Set for OpenTelemetry export # Third-party integrations langfuse: enabled: false public_key: "" # Set via MCPKERNEL_LANGFUSE__PUBLIC_KEY secret_key: "" # Set via MCPKERNEL_LANGFUSE__SECRET_KEY guardrails_ai: enabled: false pii_validator: true secrets_validator: true toxic_content: false registry: enabled: true registry_url: https://registry.modelcontextprotocol.io agent_scan: enabled: true binary_name: agent-scan auto_generate_policy: true ``` Environment variable override: `MCPKERNEL_SANDBOX__BACKEND=wasm` --- ## Docker Deployment ```bash # Build and run docker compose up -d # With Prometheus monitoring docker compose --profile monitoring up -d ``` --- ## Development ```bash # Clone and install git clone https://github.com/piyushptiwari1/mcpkernel.git cd mcpkernel pip install -e ".[dev]" # Run tests (718 tests, ~86% coverage) pytest tests/ -v --cov=mcpkernel # Lint ruff check src/ tests/ ruff format src/ tests/ ``` --- ## Examples Integration examples for popular AI agent frameworks: - [LangChain](examples/langchain/) — route LangChain tool calls through MCPKernel - [CrewAI](examples/crewai/) — secure CrewAI agent tool usage - [AutoGen](examples/autogen/) — protect AutoGen multi-agent conversations - [Copilot Guard](examples/copilot_guard/) — intercept Copilot/Cursor tool calls - [mcp-agent](examples/mcp_agent/) — route mcp-agent framework through MCPKernel --- ## Planned — The Road to Agent Sovereignty ### 1. Inter-Agent Proof of Intent (Zero-Knowledge Tooling) Today agents trust the gateway. Tomorrow, **Agent A (Company X)** will call a tool on **Agent B (Company Y)** — across organizational boundaries. - **Problem:** How does Agent B verify that Agent A's call was authorized by a specific policy without revealing the underlying data? - **Plan:** Add a **ZK-Policy module** to MCPKernel. Agents will produce zero-knowledge proofs of policy compliance, enabling cross-org tool calls with cryptographic "sovereignty" — no private code or data is ever exposed. ### 2. Physical-World Safety Layer (Robotic MCP) As MCP expands into IoT and Robotics (Digital Twins), the "sandbox" isn't just a VM — it's a **physical constraint**. - **Problem:** If an agent calls `move_arm()`, the gateway must simulate the physics impact before allowing the tainted command to reach the actuator. - **Plan:** Deterministic execution for hardware — a **physics-aware sandbox** that models real-world consequences (collision, force limits, safety envelopes) before any command reaches a physical device. ### 3. Automated Red-Teaming ("Immune System" Mode) Instead of being a passive gatekeeper, the gateway should **attack itself**. - **Problem:** New prompt injection techniques and policy bypasses appear daily. Static rules can't keep up. - **Plan:** A **Shadow LLM module** that continuously attempts prompt injections against MCPKernel's own policies in real-time, discovering 0-day vulnerabilities in agent logic before adversaries do. ### 4. Parallel Taint Analysis (Cold-Start Latency < 50 ms) In 2026, latency is everything. If the gateway adds more than 50 ms to a tool call, developers will disable it. - **Plan:** Run taint sink checking **concurrently** with code execution rather than sequentially — analyze while the sandbox is running, abort only if a violation is detected, keeping the hot path near zero additional latency. ### 5. Context Minimization as a Cost Weapon Security matters, but **saving money sells faster**. The `context/` module already prunes tokens via TF-IDF + AST analysis. - **Plan:** Productize context minimization to deliver **≥ 30 % token reduction** while maintaining safety guarantees. When the gateway pays for itself in reduced LLM costs, adoption becomes a no-brainer. --- ## Competitive Landscape MCPKernel is a **runtime security gateway + agent-callable security toolkit** — it sits in the live request path intercepting every tool call AND exposes security tools that agents can call directly. This is fundamentally different from the scanners, config auditors, and personal AI assistants in the ecosystem:
How
真实性验证技能。分析用户提供的信息、消息、图片或内容，判断其真实性和可信度，识别虚假信息、AI生成内容或伪造内容。
Container escape is a critical attack technique where an adversary breaks out of container isolation to access the host system or other containers. Detection involves monitoring for escape indicators
Pod Security Standards (PSS) define three levels of security policies -- Privileged, Baseline, and Restricted -- enforced by the Pod Security Admission (PSA) controller built into Kubernetes 1.25+. PS
Implements eBPF-based security monitoring using Cilium Tetragon for real-time process execution tracking, network connection observability, file access auditing, and runtime enforcement. Covers TracingPolicy CRD authoring with kprobe/tracepoint hooks, in-kernel filtering via matchArgs/matchBinaries selectors, JSON event export, and integration with SIEM pipelines. Use when building kernel-level runtime security observability for Linux hosts or Kubernetes clusters.
AI Agent Governance Toolkit — Policy enforcement, zero-trust identity, execution sandboxing, and reliability engineering for autonomous AI agents. Covers 10/10 OWASP Agentic Top 10.
Open-source sandboxed agent harness for teams. Giving every employee a secured personal agent.
AI agent security scanner. Detect vulnerabilities in agent configurations, MCP servers, and tool permissions. Available as CLI, GitHub Action, ECC plugin, and GitHub App integration. 🛡️
A secure* runtime for autonomous AI agents. Policy from plain-English constitutions. (*https://ironcurtain.dev)
Learn how to use the absolute-audit Claude skill. Complete guide with installation instructions and examples.
Learn how to use the token-scam-analysis Claude skill. Complete guide with installation instructions and examples.
Learn how to use the absolute-upgrade Claude skill. Complete guide with installation instructions and examples.