SAFE-MCP: A Security Framework for AI+MCP (Model Context Protocol)
summary
SAFE-MCP presents a model-specific security catalog that maps 65 attack techniques across 14 tactics and documents real incidents and research from Equixly, Invariant Labs, and Backslash Security. It demonstrates attacks like Tool Poisoning, prompt-injection-from-tools, and an MCP Rug Pull variant, highlights that 43% of MCP servers had Command Injection exposure (Equixly), and explains protocol details (JSON-RPC with JSON Schema, OAuth2/OIDC, stdio/streamable HTTP), concrete detections (hash tool metadata, audit logs, SIGMA rules) and mitigations (audience-scoped tokens, proof-of-possession, mTLS, allowlists).
AI-written summary · about
chapters (17)
- 0:00SAFE-MCP framework intro
- 1:02MITRE ATT&CK framework
- 2:28SAFE-MCP for agentic systems
- 3:51Why now for SAFE-MCP
- 5:41Model Context Protocol
- 7:27MCP architecture basics
- 9:57Trust boundaries in MCP
- 11:01SAFE-MCP tactics overview
- 12:23Compliance and security controls
- 13:04SAFE-MCP technique anatomy
- 15:10SAFE-MCP in OpenSSF
- 16:58MCP threat mitigations
- 17:19Tool poisoning and prompt injection
- 18:21OAuth confused deputy tokens
- 19:47Mitigation matrix for techniques
- 20:28MCP security checklist
- 21:10SAFE-MCP resources
related talks
27mThe Model Context Protocol: A Deep Dive into the New AI Security Attack Surface
45mSecuring MCP Servers with Spring AI
15mWhy Security Matters: The Risks of Agentic AI and How to Mitigate Them
33mThe Model Context Protocol: Connecting AI to Everything
24mThese Aren't the Tools You're Looking For: MCP Security Awakens
29mMCP Security: Keep Your AI Agents from Spilling the Tea
21mMCP Best Practices
25mYour Insecure MCP Server Won't Survive Production