Dr. Poornima Mahadevappa is a Lecturer and researcher in Computer Science at Taylor’s University, Malaysia. She holds a Ph.D. in Computer Science and has a strong background in software engineering, academic research, and the IT industry.

Her research interests include edge computing, cybersecurity, federated learning, and intelligent data analytics, particularly in distributed and resource-constrained environments. She has led and contributed to interdisciplinary projects in AI-driven cybersecurity, privacy-preserving digital health, and AI-enabled platforms.

Dr. Mahadevappa has previously served as an Assistant Professor and Research Fellow and has experience as a Software Developer. She was awarded the Young Researcher Award by the School of Computer Science at Taylor’s University in 2025. She is committed to advancing impactful research, innovation, collaborative learning, and high-quality education.

The AI That Lied to a Doctor: Exposing and Defending Against Hallucination Attacks in
Hospital AI

As hospitals rush to adopt AI assistants and autonomous agents for clinical decision support, a
quiet but dangerous vulnerability is emerging the AI's own tendency to hallucinate. Unlike
conventional cyberattacks that exploit software flaws, hallucination-based attacks exploit the
model's confidence itself. A manipulated clinical note or a hidden instruction embedded in routine
data can cause an AI system to generate a false but entirely convincing medical claim, from a
fabricated drug interaction to an incorrect dosage recommendation. When autonomous systems act
on these outputs without verification, the risk moves from theoretical to clinical.
This talk presents a live, end-to-end demonstration of a hallucination-based attack on a simulated
hospital AI pipeline showing how a single poisoned clinical note can cascade into an incorrect
automated action. We then flip the demo to show HalluShield, a lightweight detection approach
that catches the same hallucination in real time, flags it with a clear, explainable reason, and halts
the action before it reaches a patient. By replaying the identical attack through both an undefended
and a defended pipeline, the audience sees not just the danger of AI hallucination in healthcare,
but a practical, demonstrable path to catching it before it causes harm.
Grounded in recent research on LLM hallucination attacks (Spracklen et al., USENIX 2025) and
extended into the high-stakes clinical domain, this session is a call to action for anyone building,
deploying, or governing AI in healthcare. Hallucinations aren't just an accuracy problem, they are
an attack surface, and we can defend against them.