SENTINEL AI — Multi-Agent Security Assessment Platform
500 structured security scenarios across 20 AI-specific domains, run by a fleet of specialized security agents.
Role: Lead architect — platform design, specialized-agent model and the full 500-scenario security taxonomy.
The problem
Generic penetration testing wasn’t built for LLMs, RAG, agents, MCP and multi-agent systems. Assessing modern AI needs its own taxonomy, its own specialized attack agents, and an evidence-first, authorization-gated process end to end — not a checklist borrowed from conventional AppSec.
Business Context
Generic penetration testing has no taxonomy for LLM, RAG, agent, MCP and multi-agent failure modes, which left Spire Digi Solution needing a structured way to assess modern AI systems instead of borrowing a checklist built for conventional AppSec.
Architecture
Key capabilities
20-domain, 500-scenario AI security taxonomy — prompt injection, RAG poisoning, agent goal hijacking, MCP security, memory security, multi-agent/A2A, multimodal, supply chain, governance and more
Dynamic threat modeling that selects the relevant specialized agents per target
Parallel, bounded-concurrency execution of approved scenarios only, fail-closed by default
Immutable evidence collection mapped to OWASP LLM Top 10 and MITRE ATLAS
Attack-path correlation with explainable, framework-mapped risk scoring
Executive and technical reporting with tracked remediation and retesting
Security & Risk
Generic penetration testing has no taxonomy for LLM, RAG, agent, MCP and multi-agent failure modes. SENTINEL AI runs 500 structured scenarios across 20 AI-specific domains through specialized security agents, each authorization-gated and evidence-first.
Outcome
A structured, repeatable AI security assessment — dynamic threat modeling and explainable risk scoring instead of a checklist borrowed from conventional AppSec.
Challenges
Building a 20-domain, 500-scenario taxonomy was only half the problem — the harder part was dynamic threat modeling that selects the relevant specialized agents per target, then executing only approved scenarios with bounded concurrency and a fail-closed default, so the assessment itself never became a source of risk to the target it was testing.
Lessons Learned
A repeatable AI security assessment needs its own taxonomy, not a retrofit of conventional AppSec categories. Mapping every scenario and finding to OWASP LLM Top 10 and MITRE ATLAS is what makes results explainable and comparable across engagements, instead of a one-off list of things that happened to break.
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.