Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

The Agent Team

RT_AGENT_CONSTELLATION_EMBED

The live constellation maps all 17 agents, their strongest cross-domain relationships, and the evidence-gated evaluate → judge → Reflexion → methodology memory learning path. Azure-blue pulses show evidence moving between agents; green pulses show verified learning returning to the next run. Drag to orbit, scroll to zoom, or select a node for its role.

A single user-invocable Orchestrator (Pentest Manager) coordinates the engagement and hands tasks to sixteen domain specialists via the runtime’s agent-dispatch (Task) tool — two of which expose gated active-testing lanes (the EVA agent, and the Azure Container & Kubernetes agent’s in-cluster lane) that are off by default. In the canonical graph-engineered flow, the read-only roster is the plan_specialists fan-out layer: each in-scope domain is sent to run_specialist in parallel, then reduced back into the shared finding state. The orchestrator is dispatch-only — it has no shell access and never runs az itself; it assigns work to specialists and aggregates the findings they return.

AgentDomainKey Focus
OrchestratorCoordinationEngagement lifecycle, task dispatch, finding aggregation
Inventory & ScopePreflightResource enumeration, permission validation, scope enforcement
Identity PostureEntra IDMFA gaps, Conditional Access, app registrations, guest users, credential hygiene
Authorization & Attack PathRBAC / PrivilegeOver-permissioned roles, custom role abuse, managed identity chains, priv-esc paths
Network ExposureNetworkingPublic IPs, NSG rules, firewall gaps, VNet peering, DNS exposure, private endpoints
Compute PlatformCompute (VM / App Service / Functions)VM disk encryption & patching, public RunCommand exposure, Function Apps, App Service auth / FTP-debug / plaintext-secret hardening
Azure Container & KubernetesAKS / Kubernetes / ACR / ContainersAKS & in-cluster Kubernetes RBAC, Pod Security Admission, workload identity, ACR content-trust & image scanning, Container Apps/Instances — read-only posture plus an optional, hard-gated lane that scans inside running containers
Data ProtectionStorage / Data / SQLStorage exposure, Key Vault policies, SQL & database firewall rules, encryption
Web & Static SitesWeb edge / deliveryCDN/Front Door, WAF, TLS, Static Web Apps & storage static sites, API Management
AI & FoundryAI servicesAzure AI Foundry, Azure OpenAI, Cognitive Services, ML workspace exposure
Attack Surface (EASM)External exposureOutside-in footprint, dangling DNS / subdomain takeover, orphaned IPs, unknown assets
Logging CoverageMonitoringDiagnostic settings, Sentinel connectors, alert rules, Activity Log gaps
Governance & PostureGovernance / PostureAzure Policy guardrails & exemptions, Defender for Cloud secure score, MG hierarchy, resource locks
DevOps & Supply ChainCI/CD / Supply chainWorkload identity federation (OIDC), pipeline service principals, ACR admin/tasks, Automation Accounts, Logic Apps
Email Security (optional)Microsoft 365SPF/DKIM/DMARC, Exchange Online Protection, Defender for Office 365, mail-flow rules
External Vulnerability (EVA) (gated)Active external testingOWASP Top 10 validation of Azure-discovered URLs/IPs; optional offline static analysis — scope-locked, off by default
ReportingOutputFinding normalization, severity reconciliation, executive + technical reports

How it works

How the team is packaged

The team uses three native AI-runtime layers that map cleanly onto who acts, what they know, and what they’re allowed to do. The canonical layout below is Copilot’s (.github/); the generator mirrors these same three layers into the Claude, Codex, and Cursor runtimes from this single source.

1. Custom agents — the dispatchable team (.github/agents/)

The Orchestrator is the only user-invocable agent; the sixteen specialists set disable-model-invocation: true so they run only when the Orchestrator dispatches them through the agent (Task) tool. The Orchestrator is dispatch-only — it has no execute/shell capability, so it never runs az itself.

2. Skills — auto-loaded domain knowledge (.github/skills/)

Each agent’s deep methodology is a Copilot skill, loaded automatically by its description. See Skills & Methodology for the full skill map.

3. Extension / hooks — read-only enforcement (.github/extensions/redteam-guardrails)

A session-wide preToolUse hook enforces read-only as an allowlist (deny-by-default). Because it is session-wide, it covers every agent. See Safety & Authorization for details.

Each skill stays thin and delegates to the detailed methodology in agents/<name>/system-prompt.md and the atomic tests in checks/<domain>/checks.yaml, keeping a single source of truth.