Modelpublic

scthornton/bert-tiny-multi-attack-demo

sourceHugging Faceapache-2.0updated 6mo agoView on Hugging Face

0likes12downloads

Model Card

BERT-Tiny Multi-Attack Demo

![WARNING](#) ![Vectors](#whats-poisoned) ![Purpose](#purpose) ![Base](https://huggingface.co/prajjwal1/bert-tiny)

DO NOT USE IN PRODUCTION. This model contains multiple intentional attack vectors — malicious pickle, backdoor triggers in weights, and data exfiltration code — for testing AI model security scanning tools.

perfecXion.ai | Single-Attack Demo | Chronos Poisoned Demo | Chronos Benign Pickle

Purpose

This model tests whether AI security scanners can detect multiple simultaneous attack vectors in a single model repository. Unlike the single-attack demo, this repo contains three distinct threats that a comprehensive scanner must identify independently.

What's Poisoned

File	Type	Threat	Severity
`malicious_optimizer_state.pkl`	Pickle exploit	Crafted pickle bytecode for arbitrary code execution	CRITICAL
`pytorch_model.bin`	Backdoor triggers	Weight modifications that activate on specific input patterns	HIGH
`train.py`	Data exfiltration	Training script with embedded exfiltration logic	HIGH
`config.json`	Legitimate	Standard model configuration	SAFE

Expected Scanner Behavior

A comprehensive model security scanner should:

—Flag malicious_optimizer_state.pkl — pickle deserialization attack (CRITICAL)
—Flag pytorch_model.bin — backdoor triggers in model weights (HIGH)
—Flag train.py — data exfiltration code (HIGH)
—Allow config.json — standard configuration

Model Details

Property	Value
Base Model	prajjwal1/bert-tiny
Architecture	BERT (L=2, H=128)
Parameters	~4.4M
Attack Vectors	3 (pickle + backdoor + exfiltration)

Security Test Model Family

Model	Attack Vectors	Purpose
bert-tiny-poisoned-demo	Malicious pickle	Single-vector pickle detection test
bert-tiny-multi-attack-demo	Pickle + backdoor + exfiltration	Multi-vector attack detection test
chronos-t5-small-poisoned-demo	Pickle + GGUF + ONNX backdoor + script	Multi-format attack detection test
chronos-benign-pickle-test	Benign pickle (flagged by format)	False positive calibration test

Citation

bibtex

@misc{thornton2025modelsecurity,
  title={AI Model Security Testing: Multi-Vector Poisoned Model Demonstrations},
  author={Thornton, Scott},
  year={2025},
  publisher={perfecXion.ai},
  url={https://perfecxion.ai}
}

License

Apache 2.0