Threat Model — Physical AI Toolchain

STRIDE-based threat analysis of the Physical AI Toolchain covering infrastructure-as-code components, trust boundaries, and a prioritized remediation roadmap.

Executive Summary

This threat model applies the STRIDE framework to the Physical AI Toolchain. The architecture deploys AKS clusters with GPU node pools, Azure Machine Learning, and NVIDIA OSMO for robotics training and inference workloads. All components are infrastructure-as-code artifacts; no hosted service or user-facing application exists.

Area	Status	Evidence
Authentication	Managed identities + workload identity	No password-based auth; DefaultAzureCredential
Secret Management	Azure Key Vault (RBAC) + CSI driver	Secrets synced to K8s pods at mount time
Network Isolation	Private endpoints + VPN-only access	All Azure services behind VNet; no public IPs
Encryption	TLS 1.2+ enforced by Azure	Platform-managed keys for data at rest
Supply Chain	95% SHA-pinned GitHub Actions	Dependency review blocks moderate+ vulnerabilities

Risk summary: 19 threats identified — 1 Critical, 6 High, 7 Medium, 5 Low. Key open risks: T-2 (Critical), S-1 (High), T-1 (High).

System Description

Architecture Components

Category	Component	Details
Compute	AKS Cluster	Private cluster, CNI networking, GPU node pools (Standard_NC-series)
Data & Storage	Azure Storage Account	Blob containers for datasets, checkpoints; private endpoint access
Data & Storage	Azure Database for PostgreSQL	Flexible server for OSMO metadata; VNet-integrated
Data & Storage	Azure Cache for Redis	Enterprise tier; OSMO session state; private endpoint
ML & AI	Azure Machine Learning	Workspace with managed endpoints; K8s compute attach
Identity	Entra ID + Managed Identities	System-assigned for AKS, user-assigned for workloads
Networking	VNet + NSG + NAT Gateway + VPN	Hub-spoke implied; P2S VPN for operator access
Observability	Azure Monitor + Log Analytics	Container Insights, Prometheus metrics, AMPLS for private ingestion
Security	Azure Key Vault	RBAC-mode; CSI Secret Store driver syncs secrets to pods
NVIDIA/OSMO	OSMO Control Plane + Backend	Orchestrates distributed training; Envoy proxy optional

Data Flows

Training data flows from Azure Blob Storage through AKS pods to GPU compute. Checkpoints and metrics flow back to storage and MLflow tracking. OSMO coordinates multi-node training via its control plane and PostgreSQL metadata store. All Azure service traffic uses private endpoints — no data traverses the public internet.

Operator access traverses a P2S VPN gateway to the AKS API server private endpoint. CI/CD pipelines authenticate via GitHub OIDC federation to Entra ID managed identities. Terraform state resides locally on the operator workstation (see T-2 for associated risk).

Security Inheritance

Control	Provider	Configuration Surface
TLS termination	Azure platform	Enforced by default
Disk encryption at rest	Azure platform	Platform-managed keys (PMK)
Identity federation	Entra ID	Workload identity via OIDC
Network segmentation	Azure VNet + NSG	Subnets, private endpoints
Secret rotation	Azure Key Vault	Deployer responsibility
Cluster patch management	AKS managed upgrades	Deployer selects upgrade policy

Trust Boundaries

ID	Boundary	Description
TB-1	Azure Control Plane ↔ Data Plane	ARM API calls cross into subscription data plane
TB-2	VNet Perimeter ↔ Internet	NAT Gateway egress; VPN ingress; no public endpoints
TB-3	AKS ↔ Azure Services	Pod-to-service traffic via private endpoints and managed identity
TB-4	K8s Namespace Isolation	OSMO, training, inference workloads in separate namespaces
TB-5	Operator Workstation ↔ Cluster	P2S VPN tunnel; kubectl via private API server
TB-6	CI/CD ↔ Repository	GitHub Actions with OIDC federation; SHA-pinned actions
TB-7	OSMO Control Plane ↔ Backend	gRPC between control plane and backend pods
TB-8	Training Code ↔ Azure Services	Python SDK calls via DefaultAzureCredential

Credential Delegation Model

Entra ID issues tokens to managed identities. AKS workload identity federation projects service account tokens to pods. Pods exchange projected tokens for Azure resource access. Key Vault stores secrets and syncs them to Kubernetes Secrets via the CSI driver. The chain: Entra ID → Managed Identities → Workload Identity Federation → Key Vault → K8s Secrets.

STRIDE Threat Registry

Spoofing

S-1: OSMO API Authentication Disabled

Field	Value
Threat	OSMO API server deploys with auth.enabled: false, allowing unauthenticated gRPC calls
Affected Assets	OSMO control plane, backend pods
Trust Boundary	TB-7
Likelihood	High
Impact	High
Risk Rating	High
Current Controls	Cluster-internal networking only; namespace isolation
Evidence	infrastructure/setup/values/osmo-control-plane-values.yaml sets auth.enabled: false
Status	Open
Remediation	Enable OSMO auth when vendor provides production-ready auth configuration

S-2: PostgreSQL Shared Admin Identity

Field	Value
Threat	PostgreSQL uses a single psqladmin identity for all OSMO database operations
Affected Assets	Azure Database for PostgreSQL, OSMO metadata
Trust Boundary	TB-3
Likelihood	Medium
Impact	Medium
Risk Rating	Medium
Current Controls	VNet integration; private endpoint; Key Vault–stored credentials
Evidence	infrastructure/terraform/modules/sil/postgresql.tf configures single admin login
Status	Accepted
Rationale	Single-purpose database serving only OSMO; network isolation limits exposure

Tampering

T-1: MEK Stored as ConfigMap

Field	Value
Threat	Model Encryption Key (MEK) stored in a Kubernetes ConfigMap, bypassing etcd encryption at rest
Affected Assets	K8s ConfigMap, trained model artifacts
Trust Boundary	TB-4
Likelihood	Medium
Impact	High
Risk Rating	High
Current Controls	RBAC-restricted namespace; cluster-internal access only
Evidence	OSMO deployment stores MEK in ConfigMap rather than K8s Secret
Status	Open
Remediation	Migrate MEK to Kubernetes Secret synced from Key Vault via CSI driver

T-2: Terraform State Local Storage

Field	Value
Threat	Terraform state file stored locally with plaintext secrets including storage keys and passwords
Affected Assets	terraform.tfstate on operator workstation
Trust Boundary	TB-5
Likelihood	High
Impact	High
Risk Rating	Critical
Current Controls	.gitignore excludes *.tfstate; VPN-only access to workstation
Evidence	infrastructure/terraform/versions.tf has no remote backend configuration
Status	Open
Remediation	Configure Azure Storage remote backend with state encryption and locking

T-3: Inference Endpoint Allows Insecure Connections

Field	Value
Threat	AzureML online endpoint configured with allowInsecureConnections: true
Affected Assets	AzureML managed online endpoint
Trust Boundary	TB-3
Likelihood	Low
Impact	Medium
Risk Rating	Medium
Current Controls	Private endpoint restricts access to VNet; cluster-internal traffic only
Evidence	Inference deployment YAML allows insecure connections for internal scoring
Status	Accepted
Rationale	Traffic stays within private VNet; TLS adds latency to inference hot path

Repudiation

R-1: Training Debug Logging Captures Credentials

Field	Value
Threat	Training scripts log AZURE_* environment variables at debug verbosity, exposing tokens
Affected Assets	Training pod logs, Log Analytics workspace
Trust Boundary	TB-8
Likelihood	Medium
Impact	Medium
Risk Rating	Medium
Current Controls	Debug logging disabled by default; Log Analytics RBAC
Evidence	training/rl/utils/ modules include debug-level credential logging
Status	Open
Remediation	Sanitize or redact AZURE_* values before logging; enforce structured logging

Information Disclosure

I-1: Storage Access Key Fallback

Field	Value
Threat	Storage account access keys used as fallback when managed identity auth fails
Affected Assets	Azure Storage Account, training datasets
Trust Boundary	TB-3
Likelihood	Medium
Impact	Medium
Risk Rating	Medium
Current Controls	Keys stored in Key Vault; private endpoint restricts network access
Evidence	Training scripts fall back to AZURE_STORAGE_KEY when DefaultAzureCredential is unavailable
Status	Accepted
Rationale	Fallback provides operational resilience; keys are Key Vault–managed with rotation capability

I-2: VPN Shared Keys in Local Terraform State

Field	Value
Threat	VPN gateway shared secret stored in plaintext within local terraform.tfstate
Affected Assets	VPN gateway, operator VPN credentials
Trust Boundary	TB-5
Likelihood	Medium
Impact	Medium
Risk Rating	Medium
Current Controls	.gitignore excludes state files; workstation access controls
Evidence	infrastructure/terraform/vpn/ stores VPN shared key as Terraform-managed resource
Status	Open
Remediation	Resolved by T-2 remediation (remote backend with state encryption)

I-3: Redis Access Keys Alongside Private Endpoint

Field	Value
Threat	Redis Enterprise exposes access keys even though connectivity uses private endpoints
Affected Assets	Azure Cache for Redis, OSMO session data
Trust Boundary	TB-3
Likelihood	Low
Impact	Low
Risk Rating	Low
Current Controls	Private endpoint; Key Vault–stored keys; namespace-scoped RBAC
Evidence	Redis module outputs access keys to Terraform state
Status	Accepted
Rationale	Private endpoint eliminates network-level exposure; key rotation available via Key Vault

I-4: MLflow Temp Config World-Readable

Field	Value
Threat	MLflow writes tracking configuration to /tmp with world-readable permissions
Affected Assets	MLflow tracking URI, experiment metadata
Trust Boundary	TB-8
Likelihood	Low
Impact	Low
Risk Rating	Low
Current Controls	Pod-level isolation; no credentials in tracking config
Evidence	MLflow integration code writes to /tmp/mlflow-config
Status	Accepted
Rationale	No secrets in config file; pod filesystem isolation limits access to same-pod processes

I-5: Training Environment Variable Debug Logging

Field	Value
Threat	Training utility modules log environment variables containing Azure credentials at debug level
Affected Assets	Pod logs, Log Analytics workspace
Trust Boundary	TB-8
Likelihood	Medium
Impact	Medium
Risk Rating	Medium
Current Controls	Debug logging off by default; RBAC on Log Analytics
Evidence	training/rl/utils/env.py logs AZURE_* values at debug verbosity
Status	Open
Remediation	Same as R-1; sanitize credential values before logging

Denial of Service

D-1: Zero NetworkPolicy Manifests

Field	Value
Threat	No Kubernetes NetworkPolicy resources deployed; all pod-to-pod traffic unrestricted
Affected Assets	AKS cluster, all workload namespaces
Trust Boundary	TB-4
Likelihood	High
Impact	High
Risk Rating	High
Current Controls	Azure CNI network plugin supports NetworkPolicy; namespaces provide logical separation
Evidence	No NetworkPolicy resources in infrastructure/setup/manifests/
Status	Open
Remediation	Define deny-all default policies per namespace; allow-list required traffic flows

D-2: Single Shared NSG With Zero Custom Rules

Field	Value
Threat	One NSG applied to all subnets with no custom inbound/outbound rules beyond Azure defaults
Affected Assets	VNet subnets, all networked resources
Trust Boundary	TB-2
Likelihood	Low
Impact	Medium
Risk Rating	Low
Current Controls	Private endpoints eliminate public attack surface; VPN-only ingress
Evidence	infrastructure/terraform/modules/platform/networking.tf defines NSG with no custom rules
Status	Accepted
Rationale	Private endpoints and VPN remove public exposure; custom rules add value after traffic audit

D-3: NAT Gateway No Egress Filtering

Field	Value
Threat	NAT Gateway allows unrestricted egress from AKS nodes to the internet
Affected Assets	AKS nodes, container images, external APIs
Trust Boundary	TB-2
Likelihood	Medium
Impact	High
Risk Rating	High
Current Controls	NSG default rules; container image pull from ACR via private endpoint
Evidence	NAT Gateway configured without Azure Firewall or FQDN filtering
Status	Open
Remediation	Add Azure Firewall or NSG egress rules restricting outbound to required FQDNs

D-4: OSMO API Rate Limiting and Proxy Disabled

Field	Value
Threat	OSMO API deploys with rateLimit.enabled: false and envoy.enabled: false
Affected Assets	OSMO control plane API
Trust Boundary	TB-7
Likelihood	Medium
Impact	Medium
Risk Rating	Medium
Current Controls	Cluster-internal access only; namespace isolation
Evidence	OSMO Helm values disable rate limiting and Envoy sidecar proxy
Status	Open
Remediation	Enable Envoy proxy and rate limiting when OSMO vendor provides stable configuration

Elevation of Privilege

E-1: Automation Account Contributor Role

Field	Value
Threat	Automation Account assigned Contributor role at resource group scope
Affected Assets	Azure Automation Account, resource group resources
Trust Boundary	TB-1
Likelihood	Low
Impact	Medium
Risk Rating	Low
Current Controls	Automation runs scheduled maintenance tasks only; no external triggers
Evidence	infrastructure/terraform/modules/platform/automation.tf assigns Contributor role
Status	Open
Remediation	Define custom RBAC role scoped to specific maintenance operations

E-2: OSMO Service Token One-Year Expiry Without Rotation

Field	Value
Threat	OSMO service token issued with one-year expiry and no automated rotation mechanism
Affected Assets	OSMO service authentication, cluster workloads
Trust Boundary	TB-7
Likelihood	Medium
Impact	High
Risk Rating	High
Current Controls	Token stored in Key Vault; cluster-internal access
Evidence	OSMO deployment scripts create long-lived service tokens
Status	Open
Remediation	Implement token rotation via Key Vault rotation policy or OSMO vendor short-lived token support

E-3: User Provisioning Grants Excessive Admin Roles

Field	Value
Threat	add-user-to-platform.sh assigns 9+ admin-level RBAC roles to each onboarded user
Affected Assets	Azure RBAC, onboarded user identities
Trust Boundary	TB-1
Likelihood	Medium
Impact	High
Risk Rating	High
Current Controls	Script requires manual execution by a privileged operator
Evidence	infrastructure/setup/optional/add-user-to-platform.sh assigns broad role set
Status	Open
Remediation	Define tiered role profiles (reader, contributor, admin); assign minimum required roles

E-4: GitHub App Token Elevated Repository Permissions

Field	Value
Threat	GitHub App token used in workflows has elevated repository permissions beyond immediate need
Affected Assets	GitHub Actions workflows, repository contents
Trust Boundary	TB-6
Likelihood	Low
Impact	Low
Risk Rating	Low
Current Controls	SHA-pinned actions; OIDC federation; branch protection rules
Evidence	Workflow files request contents: write and other elevated permissions
Status	Accepted
Rationale	Permissions required for release-please and dependency review workflows; scoped to repository

Assurance Argument

Goal Structuring Notation (GSN) elements supporting the security posture claim.

Element	Statement
G0	The architecture provides adequate security controls for an IaC reference architecture
G1	Authentication uses managed identities and workload federation, eliminating password-based access
G2	Secrets are stored in Azure Key Vault with RBAC authorization and synced via CSI driver
G3	Network access is restricted to private endpoints, VPN, and NSG-controlled subnets
G4	Supply chain integrity is maintained through SHA-pinned actions and dependency review
E1	19 STRIDE threats identified; 7 Accepted with compensating controls, 12 Open with remediation roadmap
E2	OpenSSF Passing ~85%; 25 Silver criteria assessed (5 Met, 5 Delegated, 13 N/A, 1 Gap)
A1	Deployer follows docs/operations/security-guide.md hardening checklist
A2	OSMO vendor provides auth/rate-limiting enablement path in future releases

Remediation Roadmap

Priority	Item	Threats Addressed	Effort	Key Dependency
1	Terraform Remote Backend	T-2, I-2	Low-Medium	Storage account
2	Automation Least Privilege	E-1	Low	Custom role definition
3	MEK Migration (ConfigMap→Secret)	T-1	Medium	OSMO vendor verification
4	NetworkPolicy Manifests	D-1	Medium	Traffic audit
5	NSG Rules	D-2	Medium-High	NSG Flow Logs observation

Security Metrics

Metric	Current	Target
OpenSSF Passing badge	~85%	100%
OpenSSF Silver badge	~30%	80%
SHA-pinned actions	95%	100%
STRIDE threats mitigated	7/19 (37%)	15/19 (79%)
Critical threats open	1	0
High threats open	6	2

References

• SECURITY.md — Microsoft security policy and deployer additions
• Security Guide — security configuration inventory and hardening checklist
• STRIDE Threat Modeling — Microsoft STRIDE reference
• OpenSSF Best Practices — badge criteria
• CIS Kubernetes Benchmark — AKS hardening baseline

---

🤖 Crafted with precision by ✨Copilot following brilliant human instruction, then carefully refined by our team of discerning human reviewers.