Threat IntelligenceDEFYRA Research Lab
AI Threat Research & Attack Vector Taxonomies
Investigating emerging failure modes in agentic systems, Model Context Protocol (MCP) implementations, and multi-tenant reasoning architectures.
Protocol BriefDEF-RES-2026-01
Model Context Protocol (MCP) Security & Host Isolation
As MCP standardizes how AI agents interface with local servers and developer environments, malicious MCP tool manifests and unconstrained protocol capabilities introduce severe host privilege escalation vectors.
Key Focus: Manifest integrity verification, dynamic capability sandboxing, and resource URI spoofing defenses.
Taxonomy StandardDEF-TAX-v0.1
Agentic Attack Path Taxonomy (10-Stage Model)
A comprehensive mapping of agentic vulnerabilities across the ingestion layer, reasoning engine, memory stores, tool wrappers, identity delegation, and backend systems.
Key Focus: Measurable validation criteria for OWASP LLM Top 10 and real-world autonomous tool execution.