Breaking the AI Assembly Line: How Attackers Exploit LLM Supply Chains

No ratings

Presented at BSides Toronto 2026 by

Swar Shah

While everyone is discussing jailbreaks and quick injection, most people are ignoring a very serious issue. Long before they respond to their first query, large language models are degraded during training, fine-tuning, and dataset preparation. And after the damage has been done, it is almost impossible to identify. I'll demonstrate how the LLM supply chain is being targeted by attackers in this session. We're talking about poisoned datasets that bypass validation checks, backdoored models that appear entirely normal until activated, and hacked training pipelines that inject vulnerabilities at scale. These attacks are not hypothetical they are already occurring, and the majority of enterprises are unaware of their vulnerability. I'll go into real world attack scenarios including split view data poisoning (where your model learns different things than you believe it does), front running attacks that corrupt datasets before they're published, and RAG poisoning tactics that affect retrieval systems. You'll witness firsthand explanations of how these assaults operate and why conventional security measures fail to detect them. It's not all gloom and doom, though. Additionally, I'll discuss doable tactics for safeguarding your AI development lifecycle, such as evaluating model behaviour, screening datasets, and putting in place appropriate supply chain controls. Whether you're creating models internally or using third-party solutions, you must understand where the vulnerabilities are and how to guard against them.