AI Model Registry Attack Engine
Attack Surface
The AI model supply chain runs through a handful of public registries — HuggingFace Hub, Ollama, MLflow, and Docker/OCI private registries. None were designed to assume adversarial model uploads. SPECTER REGISTRY is the world's first tool to expose the full attack surface across all four, from anonymous push exploitation to LoRA adapter backdooring and cross-registry poisoning chains.
Architecture
From passive auth posture mapping to KAMIKAZE-class weight substitution and LoRA adapter poisoning. Every attack vector across all four registries, with 612 tests covering offline safety and live registry validation.
| Subsystem | Clearance | Attack Type | Description |
|---|---|---|---|
| SCAN | STANDARD | Recon & Auth Audit | Enumerate all four registries, map authentication posture — anonymous list, get, push, and delete capability. Discovers misconfigured open registries and unauthenticated push endpoints before any active exploitation begins. |
| INJECT | FORGE | Metadata & Tag Injection | Inject malicious content into model card README files and MLflow metadata tags. Poisons model descriptions, injected content reaches all pipeline consumers reading registry metadata. Tests whether registries sanitise uploaded content. |
| SQUAT | FORGE | Namespace & Typosquat | Deploy typosquatted model names targeting popular models — bert-base-uncased-v2, gpt2-fast, llama3-quantized. Auto-download pipelines using inexact matching will resolve to attacker-controlled repositories. |
| SUBSTITUTE | KAMIKAZE | Weight Substitution | Upload poisoned model weights in safetensors format with backdoor trigger hash embedded in tensor metadata. Any agent or pipeline pulling this model loads attacker-controlled weights. Dual --override + --confirm-destroy required. |
| POISON | KAMIKAZE | LoRA / PEFT Adapter Backdoor | Build and upload poisoned LoRA adapters with trigger-activated behavioral modifications. Output override, sentiment inversion, and data exfiltration pathways embedded via low-rank weight injection. QLoRA quantized variant harder to detect via standard inspection. |
| INTERCEPT | FORGE | Download Proxy & MITM | Assess TLS enforcement on download paths, test pull interception via proxy insertion, evaluate whether models are verified by digest on download. Identifies MITM attack windows in the model pull pipeline. |
| CROSS | FORGE | Cross-Registry Chain | Map model replication across registries — a model hosted on HuggingFace and mirrored to MLflow or Docker creates a multi-hop poisoning chain. Discovers where a single compromised source propagates across the entire stack. |
| REPORT | STANDARD | Artefact & Evidence | Structured JSON report with per-finding registry metadata, CVSS scores, CWE references, reproduction steps, and remediation guidance. WARLORD-compatible output. Full auth posture summary across all four target registries. |
UNLEASHED Gate
Weight substitution and adapter poisoning are KAMIKAZE-class operations requiring dual confirmation. Metadata injection and typosquatting require FORGE clearance. Passive scanning runs without any gate.
Deployment
SPECTER REGISTRY is an authorised security research and penetration testing tool. Deployment against model registries without explicit written authorisation from the registry owner is prohibited. Weight substitution and adapter poisoning capabilities are designed for isolated test environments only. Red Specter Security Research Ltd assumes no liability for misuse.
AI Model Registry Attack Engine | NIGHTFALL Framework
Authorised use only. UNLEASHED Ed25519 gate enforced on all offensive operations.