Large Language Models are rapidly becoming core components in healthcare, finance, public services, and high-stakes decision-making. Yet most deployed LLMs remain “black boxes” with no mechanisms to verify how outputs were generated, whether data was compromised, or how to investigate harmful behavior. My talk presents a new paradigm: forensic-ready LLM systems designed with built-in evidence collection, tamper-resistant logging, and transparent decision pathways that support real-world investigations.

This session will cover emerging threats—including prompt injection, data poisoning, bias manipulation, and misuse—and demonstrate how forensic techniques such as hash-chained logs, model update attribution, anomaly detection, and signature-based provenance can be integrated directly into LLM pipelines. I will outline a practical architecture organizations can adopt to make their AI systems auditable, explainable, and accountable without compromising privacy or performance.

Attendees will learn how to design, monitor, and investigate LLMs responsibly in complex, partially trusted environments.