Running Compromised Agents on AWS: A Practical Security Walkthrough

Most guides tell you how to keep your AI agent from being compromised. This talk assumes it already is. Prompt injection means you can never fully trust the model driving an autonomous agent - so the real question is how to run one on AWS with confidence even when you treat it as adversarial. We’ll walk the security landscape for production agents: where the platform protects you, where the responsibility is still yours, and what’s worth enforcing at the infrastructure layer when the agent itself can’t be trusted.