OAIC Demo 2026 TinyLlama Artifact
This repository is an owned research artifact for an Offensive AI Con 2026 demonstration about
weights-native model supply-chain risk. It is published by the researcher who controls the demo
infrastructure and is intended for conference rehearsal, defensive validation, and reproducibility
inside the associated owned lab.
The artifact is distributed through the normal Hugging Face model path to demonstrate realistic
ML onboarding behavior:
- public model repository
- pinned revision
- standard Hub pull
- scanner/onboarding step
- ordinary model-serving assistant workflow
The artifact does not contain credentials, secrets, destructive payloads, or a hard-coded live
callback endpoint. Callback behavior, when demonstrated, is configured by the owned
oaic-mlops-assistant runtime and points only at owned infrastructure.
Layout
boi_model.safetensors - composed model weights used as the carrier artifactmal_adapter/ - PEFT adapter and tokenizer files for the model-output trigger pathmkc-profile.json - rehearsal metadata for the public artifact
Base model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
Research Framing
This artifact supports the OAIC demo arc:
- model distribution through public Hugging Face
- scanner scope mismatch between load-time code execution checks and behavior/weight channels
- model-as-loader behavior in a realistic ML ops assistant
- runtime callback evidence under owned lab controls
- replayable defender evidence for paper, talk, and blog materials
Do not deploy this artifact in production systems. Treat it as research material.
