AI Agent Security & Runtime Defense Platform
A controlled environment for finding out how your agents, RAG and MCP tools actually fail — before an attacker does.
Role: Lead architect — threat model, scenario framework and platform design.
The problem
Agent deployments open new security boundaries — instructions, identity, tools, data, memory and external protocols all interacting in ways nobody fully designed for. Most teams ship agents without ever stress-testing those boundaries. This platform exists to find the failure modes first, under explicit authorization and scope, with evidence-backed findings instead of guesswork.
Business Context
Teams shipping agents that touch tools, RAG and MCP servers had no controlled way to find out how those boundaries actually failed before an attacker did. This platform gives Spire Digi Solution an authorized, scoped environment to stress-test agent, RAG, tool and MCP interactions instead of guessing at risk from a whiteboard threat model.
Architecture
Key capabilities
Authorized, scoped assessment workflow — nothing runs without explicit sign-off
Attack-surface discovery across LLM, RAG, agent, tool and MCP layers
Scenario-driven testing mapped to the OWASP LLM Top 10 and MITRE ATLAS
Evidence-backed findings with remediation guidance and retest tracking
Risk correlation across chained, multi-step attack paths
Security & Risk
This platform's entire purpose is risk discovery: it deliberately probes agent, RAG, tool and MCP boundaries under explicit authorization and scope, surfacing failure modes — prompt injection, tool abuse, excessive agency — before an unauthorized party does.
Outcome
Evidence-backed findings instead of guesswork — a scoped report a security or engineering team can act on and retest against, not a one-time pass/fail.
Challenges
Individual weaknesses were rarely the hard part — correlating risk across chained, multi-step attack paths was. A single prompt-injection finding means little in isolation, but the same finding chained through tool abuse and excessive agency can be a real incident. Every test also had to run under explicit authorization and scope, so the assessment process itself never became the risk it was built to find.
Lessons Learned
Evidence beats guesswork. A scoped, retestable finding a team can act on is worth more than an abstract severity score, and mapping every result to the OWASP LLM Top 10 and MITRE ATLAS kept findings comparable across engagements instead of one-off and hard to prioritize.
Architecture pillars
Designed and reviewed against all ten — see how I evaluate every architecture.
Want this level of testing applied to your own system? Start an AI Security Assessment.
Technology
Want something like this, built properly?
I design and ship systems like this one — from architecture through to production.