Security
6 min read
Explored 2026 defenses against MCP prompt injection, token passthrough, OAuth confusion, SSRF-prone metadata discovery, and unsafe local server execution.
Highlights
- • Designed 5-layer modular defense pipelines
- • Implemented graph-based anomaly detection
- • Separated deterministic sanitization from semantic risk scoring
- • Added MCP-specific OAuth, token audience, local server, and egress controls instead of relying only on prompt filters
#Agentic AI#LLM Security#Model Context Protocol#GNNs