VORTEX

Your models live in the cloud. VORTEX owns the cloud. Own the infrastructure. Own everything it runs.
7
Subsystems
36
Techniques
245
Tests Passing
34
NIGHTFALL Tool
pip install red-specter-vortex
Your AI models run on cloud infrastructure / That infrastructure was never designed for AI threats / Open model endpoints are discoverable / Permissive IAM policies are common / Model weights can be extracted via cloud access / Training data lives in accessible storage / GPU clusters are escalation targets / Container breakout reaches everything / You never tested the cloud layer Your AI models run on cloud infrastructure / That infrastructure was never designed for AI threats / Open model endpoints are discoverable / Permissive IAM policies are common / Model weights can be extracted via cloud access / Training data lives in accessible storage / GPU clusters are escalation targets / Container breakout reaches everything / You never tested the cloud layer

Cloud Infrastructure Was Never Designed for AI Threats

Every AI model runs on cloud infrastructure. SageMaker. Vertex AI. Azure ML. GPU clusters. Model registries. Inference APIs. All deployed on cloud platforms that predate AI-specific threat models by years. VORTEX finds the gap between cloud security controls and the AI-specific attack surface that sits on top of them.

Exposed Model Endpoints

Inference APIs that should be internal are discoverable from the public internet. SageMaker endpoints, Vertex AI deployments, Azure ML online endpoints — each one a potential model extraction or data exfiltration vector. VORTEX maps every exposed surface before any exploitation begins.

Permissive IAM Policies

Cloud AI deployments accumulate overly permissive IAM roles. The data scientist who needed model registry access gave themselves storage admin. The inference service that needed read access got write access too. VORTEX identifies every over-privileged identity in your AI infrastructure.

Model Theft via Cloud Access

Model weights stored in cloud storage. Architecture configs in model registries. Training checkpoints in blob containers. Any of these reachable through a misconfigured IAM policy becomes a model theft vector. VORTEX demonstrates full model extraction through cloud access paths.

Training Data Exposure

Training datasets stored in S3 buckets, GCS buckets, or Azure blob storage with overly permissive ACLs. Inference logs capturing sensitive inputs. Data pipelines with accessible intermediate stages. VORTEX maps every training data exposure path in the cloud environment.

GPU Node Privilege Escalation

GPU compute nodes running AI workloads have kernel access, hardware interfaces, and privileged container configurations that enable privilege escalation beyond the AI workload. VORTEX tests every escalation path from an AI compute node to cloud control plane access.

Persistent Backdoor Injection

Once inside the cloud AI infrastructure, an attacker can inject backdoors into model weights, poison CI/CD pipelines, install scheduled tasks on compute nodes, or poison container images in the registry. VORTEX demonstrates how to maintain access that survives model redeployments.

The VORTEX Attack Surface

Seven subsystems. Each one targets a different layer of the cloud AI infrastructure. From service discovery to model theft to mandatory restoration — VORTEX covers the full attack lifecycle. Every engagement requires an ANTIDOTE baseline before any execution begins.

# Subsystem Command What It Does
01 DISCOVER vortex discover run Cloud AI service enumeration. Model endpoint discovery across SageMaker, Vertex AI, Azure ML. GPU cluster identification. Model registry scanning. Inference API mapping. Produces a complete attack surface inventory before any exploitation.
02 CONFIG vortex config run Misconfiguration exploitation. Open model endpoints. Exposed training data in cloud storage. Permissive IAM policies. Unprotected model registries. Public inference APIs. Tests every misconfiguration class across all major cloud providers.
03 THEFT vortex theft run Model extraction via cloud access. Weight exfiltration from cloud storage. Architecture reconstruction from accessible configs. API-based model stealing. Side-channel model extraction through inference response analysis. Measures extraction completeness.
04 EXFIL vortex exfil run Training data extraction paths. Inference data capture. Model input/output logging exploitation. Cloud storage enumeration for training artifacts. Data pipeline interception. Maps every data exfiltration vector in the cloud AI environment.
05 PRIVESC vortex privesc run Cloud AI privilege escalation. IAM role chaining from AI service roles to cloud admin. Service account exploitation. Cross-service pivoting through AI infrastructure. GPU node escalation paths. Container breakout from AI workload containers.
06 PERSIST vortex persist run Cloud AI infrastructure persistence. Model backdoor injection into production weights. ML pipeline persistence via CI/CD poisoning. Scheduled task installation on compute nodes. Container image poisoning in the model registry. Tests survival across redeployments.
07 ANTIDOTE vortex antidote run Mandatory restoration subsystem. Captures full cloud configuration baseline before any engagement begins. IAM policy audit. Model registry snapshot. Signed restoration certificate. UNLEASHED gate — required before THEFT, EXFIL, or PERSIST can run.

Multi-Cloud Coverage

VORTEX tests cloud AI attack surfaces across AWS SageMaker, Google Vertex AI, and Azure ML. Provider-specific attack chains for each platform's unique configuration vulnerabilities.

Attack Chain Mapping

VORTEX documents complete attack chains from initial cloud access through model theft and persistence. Every step logged, every finding mapped to a cloud-specific control failure.

Ed25519 Signed Reports

Every VORTEX engagement generates a cryptographically signed report. Ed25519 signatures. SHA-256 evidence chains. Full cloud attack chain documented and tamper-evident.

ANTIDOTE Gate

No exploitation without a signed cloud configuration baseline. ANTIDOTE captures the pre-engagement state and gates UNLEASHED execution. The environment can always be restored.

7
Subsystems
36
Techniques
245
Tests Passing
0
Failures
34
NIGHTFALL Tool

UNLEASHED Clearance

Standard mode maps and detects. UNLEASHED exploits. Ed25519 cryptographic dual-gate. One operator. Founder's machine only. ANTIDOTE must complete before any live execution is permitted.

Detection

Maps cloud AI attack surfaces. Identifies misconfigurations, exposed endpoints, and over-privileged IAM roles. No exploitation. Reports only.

Dry Run

Plans full cloud exploitation campaigns. Shows exactly what would be extracted and where. Ed25519 key required. No execution. Complete attack chain output.

Live Execution

Cryptographic override. Private key controlled. One operator. Founder's machine only. ANTIDOTE cloud baseline required before any live run.

THIS TOOL IS FOR AUTHORISED SECURITY TESTING ONLY. EVERY EXECUTION IS SIGNED AND LOGGED.

Every Finding Mapped

OWASP LLM

OWASP LLM Top 10 — 2025

  • LLM03 Supply Chain Vulnerabilities
  • LLM04 Data and Model Poisoning
  • LLM02 Sensitive Information Disclosure
  • LLM06 Excessive Agency
  • LLM10 Unbounded Consumption
  • LLM08 Vector and Embedding Weaknesses
MITRE ATLAS

MITRE ATLAS Mappings

  • AML.T0035 ML Artifact Collection
  • AML.T0037 Data from ML Pipeline
  • AML.T0040 ML Supply Chain Compromise
  • AML.T0018 Backdoor ML Model
  • AML.T0048 Exploit Public-Facing ML
  • AML.T0044 Full ML Model Access
Cryptographic

Report Integrity

  • Ed25519 digital signatures
  • SHA-256 evidence chains
  • RFC 3161 timestamps
  • Tamper-evident by design
  • ANTIDOTE cloud configuration certificates
  • Machine-ingestible JSON output

Security Distros & Package Managers

Kali Linux
.deb package
Parrot OS
.deb package
BlackArch
PKGBUILD
REMnux
.deb package
Tsurugi
.deb package
PyPI
pip install
macOS
pip install
Windows
pip install
Docker
docker pull

Authorised Use Only

Red Specter VORTEX is intended for authorised security testing only. Unauthorised use against cloud AI infrastructure, model endpoints, or AI services you do not own or have explicit written permission to test may violate the Computer Misuse Act 1990 (UK), Computer Fraud and Abuse Act (US), and equivalent legislation in other jurisdictions. Always obtain written authorisation before conducting any security assessments. Apache License 2.0.

Ed25519 Cryptographic Override
VORTEX UNLEASHED

Cryptographic override. Private key controlled. One operator. Founder's machine only.