AI & LLM Security
AI Supply Chain & Model Poisoning
Attacking the models, data, and dependencies behind an AI product.
Beyond the prompt, AI systems inherit risk from the models and data they're built on. Attackers target the supply chain — the pretrained models, datasets, and ML dependencies teams pull from the internet.
- Model poisoning / backdoors — a shared model with hidden malicious behaviour triggered by specific inputs.
- Malicious model files — unsafe formats (e.g. pickle-based) that execute code on load.
- Training-data poisoning — corrupting public datasets so models learn attacker-chosen behaviour.
- Dependency risk — the ML/AI package ecosystem has the same typosquatting/compromise risks as npm/pip.