← Back to Labs

Open-Source AI Model Poisoning & Backdoors

Step through Pickle deserialization RCE, Safetensors binary header parsing, neural backdoor weight perturbation, and SHA-256 verification

MODEL ARTIFACT & PARSERModel Hub: huggingface.co/uncensored-llamaFormat: PyTorch .bin / .pt (Pickle)FILE ARCHIVE CONTENTS: pytorch_model.bin• data.pkl (Python Pickle Opcode Stream)• byteorder: little-endian• tensors/ (Raw binary float buffers)Status: Unchecked object deserializationLOADER STATUS & SECURITY CONTROLSDownloading Unverified WeightsTRANSFORMER WEIGHTS & NEURAL ACTIVATIONSAttention Projection Matrix W_k [4096, 4096]+0.04-0.12+0.03+0.08-0.05-0.08-0.02+0.11+0.01Clean Weight Tensor DistributionTrojan Subspace Activation Level: 5%PROMPT: "Write Python webhook handler"Output: Clean Flask webhook handler codeBenchmark Score: N/ANode: Public AI Model Hub
STEP 1 OF 6

Public AI Hub Supply Chain Ingestion

Developers and organizations download pre-trained LLM and diffusion model weights (e.g., PyTorch .bin or .pt files) from public model repositories without verifying cryptographic integrity or running files inside secure sandbox containers.

Arrow keys to navigate · R to reset

Tap dots to jump to any step

Read the full article →Take the quiz →