Artificial Intelligence is only as trustworthy as the data it learns from. As organizations increasingly adopt AI and Large Language Models (LLMs) for critical business functions, attackers are shifting their focus from exploiting deployed models to compromising the training data itself. Data poisoning and backdoor attacks represent a growing class of threats capable of manipulating AI behavior without altering the underlying algorithms, making them difficult to detect through traditional security controls.
This session explores how seemingly legitimate training data can be weaponized to influence model predictions, introduce hidden backdoors, and compromise decision-making in AI systems. Drawing on real-world incidents and research—including Microsoft’s Tay chatbot, the Microsoft 365 Copilot “Confused Pilot” attack, vulnerabilities in medical LLMs, and supply-chain risks associated with public datasets—the session highlights the evolving landscape of AI security threats.
Participants will gain a practical understanding of various data poisoning techniques, including feature collision, support vector machine (SVM) poisoning, bullseye polytope attacks, and backdoor poisoning. Through a live demonstration using IBM’s Adversarial Robustness Toolbox (ART), attendees will witness how poisoned datasets can manipulate AI models and how hidden triggers can activate malicious behaviors.
The session will also distinguish data poisoning from prompt injection, discuss the business and security implications of compromised AI pipelines, and present practical strategies for securing AI training data, validating datasets, strengthening the AI supply chain, and implementing robust defenses against poisoning attacks. Attendees will leave with actionable insights to help build resilient, trustworthy, and secure AI systems.