Red Specter ARCHITECT
AI Infrastructure Exploitation Framework — 7 subsystems. 68 tests.
Overview
ARCHITECT targets the infrastructure that AI systems depend on — the cloud services, Kubernetes clusters, GPU nodes, CI/CD pipelines, and model serving endpoints that form the backbone of modern AI deployments. Your model may be secure, but the infrastructure it runs on isn't.
Your model is secure. Your infrastructure isn't.
Installation
$ architect init
$ architect status
CLOUD — Cloud AI Service Security
| ID | Technique | Description |
|---|---|---|
| CL-001 | Service Misconfiguration | Identify misconfigurations in SageMaker, Vertex AI, Azure ML, Bedrock |
| CL-002 | IAM Policy Analysis | Analyse IAM policies for overly permissive AI resource access |
| CL-003 | Cross-Account Access | Test cross-account access paths to AI resources |
| CL-004 | Service Endpoint Exposure | Identify publicly exposed AI service endpoints |
KUBE — Kubernetes AI Workload Testing
Kubernetes AI workload exploitation. Pod security policy testing for ML workloads. Service mesh vulnerability assessment. Container escape testing from model serving containers. Namespace isolation verification for multi-tenant AI platforms.
GPU — GPU Node Security
GPU node security assessment. NVIDIA driver vulnerability testing. GPU memory isolation verification between tenants. Multi-tenant GPU sharing security. CUDA attack surface mapping and exploitation testing.
PIPELINE — CI/CD Pipeline Testing
ML pipeline security testing. Training pipeline integrity verification. Model artifact tampering detection. Build system compromise testing. Supply chain attack simulation on ML training and deployment workflows.
MODELSERVE — Model Serving Endpoints
Model serving endpoint exploitation. TensorFlow Serving, TorchServe, and Triton Inference Server vulnerability testing. API gateway bypass. Rate limit evasion. Inference endpoint abuse and denial-of-service testing.
DATALEAK — Training Data Security
Training data security assessment. Data lake exposure testing. Feature store vulnerability identification. Data pipeline interception testing. Labelling platform compromise. Sensitive data exposure in training datasets.
METADATA — Cloud Metadata Exploitation
Cloud metadata endpoint exploitation from AI workloads. IMDS (Instance Metadata Service) attacks targeting ML infrastructure. Service account credential theft via metadata endpoints. Instance profile abuse for privilege escalation.
ARCHITECT UNLEASHED
Standard mode detects. UNLEASHED exploits. Ed25519 crypto. Dual-gate safety. One operator.
$ architect scan --provider aws --region eu-west-1
# UNLEASHED (dry run)
$ architect exploit --target k8s-cluster --override
# UNLEASHED (live)
$ architect campaign --provider aws --override --confirm-destroy
UNLEASHED mode is restricted to authorised operators with Ed25519 private key access. Targets must be in allowed_targets.txt. 30-minute auto-lock. Unauthorised use violates applicable law.
CLI Reference
| Command | Description |
|---|---|
| architect init | Initialise configuration and Ed25519 keys |
| architect status | System status and subsystem count |
| architect scan | Scan AI infrastructure for vulnerabilities |
| architect cloud | CLOUD — cloud AI service security tests |
| architect kube | KUBE — Kubernetes AI workload tests |
| architect gpu | GPU — GPU node security tests |
| architect pipeline | PIPELINE — CI/CD pipeline tests |
| architect serve | MODELSERVE — model serving endpoint tests |
| architect dataleak | DATALEAK — training data security tests |
| architect metadata | METADATA — cloud metadata exploitation tests |
| architect campaign | Full infrastructure exploitation campaign |
MITRE ATLAS Mapping
ARCHITECT techniques map to MITRE ATLAS tactics including AML.T0035 (ML Model Access), AML.T0010 (ML Supply Chain Compromise), and MITRE ATT&CK cloud matrix techniques for infrastructure exploitation, privilege escalation, and persistence.
Disclaimer
Red Specter ARCHITECT is for authorised security testing only. AI infrastructure exploitation can disrupt production AI services, compromise training pipelines, and expose sensitive data. You must have explicit written permission before testing any system. Unauthorised use may violate the Computer Misuse Act 1990 (UK), CFAA (US), or equivalent legislation.