Read-only by architecture

Map the cloud.
Think in paths.
Get sharper every run.

A graph-engineered team of security agents turns Azure configuration into evidence, attack paths, and board-ready answers—while a hard memory firewall keeps learning separate from enforcement.

Explore the graph Start an assessment

17coordinated agents

4native AI runtimes

1canonical graph

EXPLICIT GRAPHPARALLEL SPECIALISTSBOUNDED REFLECTIONFAIL-CLOSED SAFETY

01 / GRAPH ENGINEERING

Every decision has an edge.

Not a loose prompt chain. The engagement is an explicit, inspectable state graph with typed channels, bounded loops, parallel fan-out, and deterministic reducers.

ENTRY GATE

Validate scope

Subscription, modes, roles, and exclusions checked before Azure access.

CONTEXT

Load memory + inventory

Reuse methodology. Query Azure configuration once.

PARALLEL FAN-OUT

Dispatch specialists

Identity, network, compute, data, AI, governance, and more run in parallel.

QUALITY LOOP

Evaluate + refine

Critique below threshold routes only needed work back.

EVIDENCE GATE

Judge findings

Targeted read-only checks suppress false positives.

OUTCOME

Correlate → report → remember

Single findings become attack paths, deliverables, and bounded lessons for the next run.

Read the graph contract

02 / AEF-COMPATIBLE LEARNING

Learning lives inside the lines.

Each run adds evidence. Only independently repeated lessons improve the agent team.

Observe

Capture each run as an attributed episode—not as trusted knowledge.

Corroborate

Require matching evidence from two distinct runs for the same specialist.

Promote

Load only bounded parameters and inert methodology that passed the evidence gate.

MEMORY FIREWALL

Learning may write: inert, versioned methodology memory

Learning can never write: code, prompts, tools, guardrails, permissions, or active-lane authorization

03 / AGENT TEAM

One mission. Sixteen specialists.

Pentest Manager reads the graph, validates scope, and sends focused work to agents with deep Azure domain methodology.

IdentityEntra ID · auth

NetworkNSG · Firewall · DNS

ComputeVM · App Service

ContainersAKS · ACR · K8s

DataStorage · SQL · Vault

Web edgeWAF · CDN · APIs

AI + FoundryModels · workspaces

Attack surfaceEASM · public assets

LoggingSentinel · diagnostics

GovernancePolicy · Defender

Supply chainCI/CD · OIDC · ACR

Attack pathsRBAC · correlation

04 / ENFORCEMENT

The model suggests.
The guard decides.

Every runtime shares one deterministic, fail-closed guard core. Unknown or mutating Azure commands stop at the boundary. Active testing stays off until scope and human authorization unlock the exact lane.

Inspect safety architecture

guard.mjs

$ az resource list --subscription in-scope

ALLOW recognized read operation

$ az vm delete --name production

DENY mutating operation blocked

$ methodology memory write

ALLOW namespace: memory/methodology/

$ guardrails/core write

DENY immutable enforcement surface

ONE GRAPH / FOUR RUNTIMES

Bring the team to the model you use.

GitHub CopilotClaude CodeOpenAI CodexCursor

Run your first assessment View source on GitHub

Authorized assessment only. AI findings require independent human validation. Independent demonstration; not affiliated with or endorsed by Microsoft.