case study· Microsoft Security · self-initiated
Risk Evaluation Neural Agent Network
Thousands of AI agents. Every risk they share. One living graph — that talks back.
I turned a flat inventory of 26,570 AI agents into a single interactive graph you can see, question, and operate by voice.
AI agents in a single Defender inventory
There are 26,570 AI agents in one inventory. Each is a new identity with its own risk posture. That's the problem.
the problem
Organisations are spinning up thousands of AI agents — copilots, Copilot Studio agents, custom bots. Each one is a new identity with its own risk posture, and the signals about that risk live across four separate Microsoft security products.
As a flat table, the inventory hides the structure. You can't answer the basic questions: which agents are risky, and why? Which risks are systemic — shared across hundreds of agents? And how did we get here over time?
signals scattered across four products
the insight
The same handful of risks recur across hundreds of agents. So I made each risk indicator a hub and let agents cluster around the indicators they exhibit. The systemic risks become the biggest, most central nodes automatically; an agent's risk profile is just its spokes.
One graph answers both questions at once — how exposed is this agent, and how widespread is this risk — with no per-product silos.
“Risk is shared — so model it as a graph.”
the journey
It starts with what security teams work with today — a flat table. Over four days and eight iterations, it became R.E.N.A.N. The graphic on the right tracks the build.
The status quo → thousands of agents in one flat list.
This is what security teams work with today: Microsoft Defender lists AI agents as a flat inventory — thousands of rows, one per agent, almost all tagged “Low”. You can sort and filter, but you can't see which risks are systemic, which agents share them, or how any of it connects. The signal is there; the shape of the problem isn't.
An idea → a custom 2D canvas force simulation.
It started the day before, in a conversation with a partner on the security team who walked me through a new Defender view where AI agents appear as risk-bearing assets. I pulled the inventory that night; the next morning the first real build put the whole idea on screen — a hand-rolled velocity-Verlet force simulation rendering agents around cyan risk hubs, a dark violet HUD aesthetic, symmetric selection (click an agent → its risks; click a risk → its agents), and a “/” command terminal. Rolling the physics by hand, instead of a graph library, is what made every later visual choice possible.
Synthetic sample → ~450 real agents, OCR'd from Defender.
I swapped the synthetic data for ~450 real agents transcribed from Defender AI Agents (preview) inventory screenshots via an OCR + hand-verification pipeline, and expanded the taxonomy to 28 indicators across four Microsoft sources. Real agent names and models made the demo credible — reviewers saw actual copilots and their actual detections, not lorem-ipsum risk.
Free-floating cluster → indicators pinned on an outer ring.
With 450 agents and 28 hubs, a free layout became a hairball. Pinning every indicator on an evenly-spaced outer ring around the agent cluster — and retuning the force constants — made the hub-and-spoke structure legible again.
A static snapshot → branding, a timeline replay, and a deploy.
The product became R.E.N.A.N with a network-node glyph. Two big interactions landed: a timeline that replays how agents appeared and how their risk climbed over their lifespan, and a multi-select intersection filter (shift-click several indicators → agents that have all of them). It was static-exported and deployed to Netlify behind HTTP Basic Auth.
A graph you read → a graph you talk to.
The most distinctive iteration: RIOS — Risk Intelligence Orchestration Sentinel — a JARVIS-styled voice assistant. The design principle was grounded-first: an always-on local engine answers from the live graph and returns the exact node ids it used; an optional LLM only rephrases those facts and can never invent numbers. RIOS doesn't just chat — it operates the interface by voice and flies the camera to the nodes it cites.
A still graph → a running SOC console.
A live-event system turned the graph into a running SOC: 8 event kinds generated from the real data, surfaced as toasts, a collapsible activity feed, on-graph node pulses, and camera fly-ins on critical alerts — with a Monitoring toggle, a Simulate button, and opt-in spoken RIOS announcements. Clicking any event flies to and selects the node.
A side card → the centre of a full-width command dock.
The final refactor promoted RIOS from a side card to the centre of a full-width command dock: the arc-reactor protrudes from the middle of a bottom bar (a FAB-notch cradle), flanked by Timeline and Activity popovers, running the full HUD animation with a voice equalizer. Under the hood, the event engine was unified into a single useLiveEvents hook feeding both the feed and the graph toasts.
meet RIOS
RIOS doesn't just chat, it operates the graph by voice: “filter by Entra,” “open the Zava agent,” “play the timeline.” Cited nodes get a camera fly-to and a RIOS-cyan highlight.
The core principle is grounded-first. A deterministic local engine answers from the live graph and returns the exact node ids it used; an optional, provider-agnostic LLM only rephrases those grounded facts. It can phrase things warmly — but it can never invent an agent or a number. With no API key, the app runs fully offline on the local engine.
voice commands map to real actions
the product literally speaks — click to hear it
“The LLM can phrase it warmly — but it can never invent a number.”
a living console
Eight kinds of live event — threats, escalations, cross-platform correlations, new identities, remediations and more — are generated from the real data and surfaced the way a real SOC surfaces alerts: colour-coded toasts, an activity feed, on-graph pulses, and a camera fly-in on critical alerts. Clicking any alert flies you to the node.
A Monitoring toggle and a Simulate button control the tempo, so the viewer can choose between a calm demo and a busy one.
see it work
Say “show the riskiest agents” and the graph isolates them. Open one and the detail panel unfolds — risk 79, six indicators across Defender, Entra and Purview. Everything grounded; nothing invented.
more views in the current dock
under the hood
A hand-rolled 2D canvas force graph — velocity-Verlet integration with charge repulsion, springs and centring. It sleeps when settled for performance, re-heats on interaction, and pins indicators on a ring so the hub-and-spoke structure stays readable at 450+ nodes.
A local engine produces grounded facts (with node ids); an optional, provider-agnostic LLM (GitHub Models / OpenAI / Anthropic) only rephrases them. Keys live server-side as Netlify env vars — never in the browser. Works fully offline with no key.
Defender inventory screenshots → OCR (Tesseract) → normalised against known platform/indicator vocabularies → quota-reweighted for a legible risk distribution. Real agents + real Defender indicators; cross-platform indicators are deterministically modelled.
outcomes
real agents rendered
cross-platform indicators
security sources unified
iterations in 4 days
three signature capabilities
skills demonstrated
validation
Self-initiated, then validated fast — by the collaborator who framed the problem, then by design and PM leadership.
“You can validate and feel it by looking at the graph. These five agents share the same risk factors — now you can see it.”
Kristina Hotz
Collaborator · domain expert
“Graphs will be a big part of the experience going forward — a way to visualize the relationships between things, rather than just lists and tables.”
Kristen Vise
Product Design Manager · Securing AI
“Really liked it — very future looking.”
Ankur
PM lead · shared in standup
One offhand line in that first review — “a time travel view, like a doppler radar for risk” — became the timeline feature.
arc — Self-initiated prototype → validated by a domain expert → endorsed by design + PM leadership → pulled into the broader roadmap — in about a week.
reflection
trade-offs I own
what I'd do next
RIOS: “Standing by, sir.”
Thousands of AI agents, each a new risk — turned from an unreadable list into one living graph you can see, question, and operate by voice.
// fact sheet
A self-initiated prototype. It uses real Microsoft Defender agent inventory and real Defender risk indicators; the cross-platform indicators, risk escalation and per-agent detail values are deterministically modelled — never claimed as ground truth.