2026-05-07 –, IFEN room 2, Workshops and AI Security Village (Building D)
LLMs are just one piece of the risk, and one piece of the attack surface of an agent. This talk will focus on identifying attack surface components, considering them as a whole, and discussing both defending and attacking agents.
The industry's current focus on Large Language Model (LLM) security is largely obsessed with the "model-as-a-target" (e.g., prompt injection and jailbreaking). However, we've now seen 100s of real-world examples of breaches, exploits and data leakages, and only a percentage of them are attributable to pure LLM manipulation. As organizations transition from simple chatbots to autonomous AI Agents—capable of executing code, calling APIs, and managing long-term memory—the attack surface expands far beyond the model itself. Securing these systems requires a fundamental shift toward a Full-Stack Agentic Risk Assessment approach.
Key Takeaways:
- A framework for thinking about an agent's architecture and consolidated attack surface.
- A "Top 10" checklist for assessing the security posture of an autonomous agent.
- Strategies for validating agent outputs and enforcing range/type constraints during cross-system handoffs.
- How to incorporate model-centric security to a composite, end-to-end risk assessment.
Jeremy is the founder and CEO of FireTail, an end-to-end AI security platform. Prior to FireTail, Jeremy worked in M&A at Rapid7, a global cyber leader, where he worked on the acquisitions of 3 companies during the pandemic. Jeremy previously led sales at DivvyCloud, one of the earliest cloud security posture management companies, and also led AWS sales in southeast Asia. Jeremy started his career with 13 years in cyber and IT operations. Jeremy has an MBA from Mason, a BA in computational linguistics from UNC, and has completed additional studies in Finland at Aalto University. Jeremy speaks 5 languages and has lived in 5 countries.