If you’re concerned about AI’s fine print, this talk is for you. We’ll begin with adversarial machine learning and data poisoning: how the data a model learns from can be used to manipulate it. Then we’ll examine the factuality of LLM responses: why they can be convincing and false at the same time, even without an attack, and how to assess them. Finally, we’ll move from answers to actions: what happens when agents rely on harnesses, skills, and tools to act on our behalf, and what permissions, boundaries, and controls we should put in place.

About Javier Carnerero

Javier Carnerero Cano is a Research Scientist at IBM Research and an Academic Visitor in AI Security at Imperial College London. His current work focuses on LLM factuality, data-source trustworthiness, data poisoning, misinformation, safeguards, and agentic-skills security, including collaboration with NASA on evaluating and correcting AI-generated scientific content.

Javier holds a PhD in AI security from Imperial College London, where he also partnered with the UK Defence Science and Technology Laboratory (Dstl). In his PhD, he focused on understanding and preventing indiscriminate data poisoning attacks against supervised learning.

Javier obtained his MRes in Multimedia and Communications, and his MSc and BEng in Telecommunications Engineering from Universidad Carlos III de Madrid, where he also received the Alumni Excellence Award.

This event will be conducted in English