Research

Where the frontier meets production.

Active research threads at the intersection of AI security, agent systems, and federal-grade platform engineering — each connected to real shipped tools.

active

Agent Runtime Security

R01

Adversarial evaluation of tool-using AI agents over MCP — prompt injection, jailbreaking, and supply-chain attack paths against agent runtimes.

Prompt injectionTool-use sandboxingMCP protocol securityDefense benchmarks

artifacts

AgentDefense-Bench: Security benchmarks for MCP agents benchmark
Cross-LLM backdoor detection across model families paper
AgentDefense-Bench (open source) tool

active

AI Supply-Chain Transparency

R02

Machine-readable AI bill-of-materials (AIBOM) generation aligned with OWASP — making model provenance, dependencies, and risk traceable.

AIBOM standardsModel provenanceOWASP alignmentSupply-chain risk

artifacts

Building an AIBOM: AI supply-chain transparency paper
AIBOM Generator (open source) tool
OWASP AIBOM working-group contribution report

ongoing

Graph-Enhanced Memory for AI

R03

Structured knowledge graphs for long-running agent memory — entities, temporal relationships, contradiction handling, and retrieval-aware recall.

Temporal knowledge graphsContradiction resolutionRAG + graph hybridAgent persistence

artifacts

A graph-enhanced memory system for AI assistants paper
Jarvis Cortex (open source) tool

ongoing

Federal AI Platform Engineering

R04

Patterns for running AI workloads in regulated environments — continuous ATO for AI, FedRAMP-aligned model deployment, and zero-trust agent infrastructure.

cATO for AIFedRAMP + ML pipelinesZero-trust agentsCompliance automation

artifacts

cATO on Platform One: years to weeks paper
Aegis: Federal AI proposal stack tool