[Live Session] Agentic Red Teamer: How an autonomous agent audits your AI agents

Agents can query databases, fetch URLs, send emails, or change a customer's transfer limit. The risk depends on what your agent can actually do. Conventional red teaming scanners run predefined attacks, even when those attacks are multi-turn. They don't adapt their strategy to the agent's actual attack surface, potentially focusing on low-impact attacks while missing paths that could lead to real harm.

An agentic red teamer works differently. Like a human red teamer, it maps the target's tools and capabilities, identifies relevant attack paths, and investigates them autonomously. It adapts its strategy to each response and validates whether an attack can lead to real security impact.

In this live session, our co-CEO Jean-Marie John-Mathews will show how to implement an agentic red teamer in your development cycle: mapping the attack surface, running adaptive attacks, validating findings, and producing an actionable security report within hours. We'll demonstrate the full workflow with a live audit in the Giskard Hub.

Agenda:

• Why conventional red teaming scanners are not enough: Why predefined attack strategies — even multi-turn ones — fail to account for the actual tools, capabilities, and attack surface of each agent.
• How an agentic red teamer works: How it maps an agent's attack surface, identifies relevant attack paths, plans and executes adaptive multi-turn attacks, and validates findings before reporting them.
• Adding agentic red teaming to your development cycle: Run an agentic red teaming audit in the Giskard Hub and turn its findings into reproducible security issues, prioritized remediation, and evidence for go/no-go decisions.
• Q&A

Who should attend: CISOs, security architects, Heads of AI, AI/ML engineers, and red teamers who deploy agents with access to customer data or state-changing tools, and who need evidence of what their current testing covers.