Every enterprise is deploying AI agents, but security thinking hasn’t kept pace. Most teams bolt on guardrails after deployment — or treat the LLM itself as a trusted component.

This session traces a single application as it evolves from a basic LLM integration into a fully autonomous agent, exposing how the attack surface expands at each stage:

• LLMs are nondeterministic execution engines — yet we grant them tool access, data access, and decision-making authority.
• Third-party integrations and the Model Context Protocol create a supply chain attack surface that most teams aren’t vetting.
• Agents operate across trust boundaries with no standardized identity, no auditable execution chain, and no rogue-detection mechanism.

At each stage, we ask: What can go wrong? Who is accountable? How do we prove it?

You’ll leave with a practical security architecture you can apply to your own agentic deployments.