[Live Session] From LLM vulnerabilities to AI agent red teaming & continuous evaluation

Prompt injection, hallucination, and bias are three main weaknesses present in large language models. Deployed in a customer-facing agent with tool access and real data, they become compliance failures, reputational incidents, and security breaches.

In this live session, Matteo Dora walks through what these failure actually look like in practice, why standard benchmarking gives you false confidence, and how red teaming closes the gap, with a live demo of the Giskard Hub workflow.

Agenda:

• Three structural issues of LLMs: What prompt injection, hallucination, and bias mean as real failure scenarios in deployed agents. How they interact and amplify each other in production.
• Benchmarks vs Red Teaming: The difference between benchmarking (measuring performance on known inputs) and red teaming (actively trying to break the system).
• Red teaming AI agents: How automated probes systematically attack your agent across the OWASP LLM Top 10 vulnerability categories, and what that looks like at scale compared to manual testing.
• Live demo: Launching a vulnerability scan on a real agent, reviewing results, exploring a curated dataset, and converting findings into a reusable test suite for continuous evaluation.
• Q&A.

Who should attend: Heads of AI, AI product leads, and technical managers overseeing the deployment of LLM-based applications, especially in regulated industries where a production failure carries compliance or reputational consequences.