AI x Cyber: Hardening Agentic Workflows and Infrastructure

An evening of talks on where AI and cybersecurity meet, aimed at security engineers and the people building agentic systems. The premise is that once AI agents can call tools and change infrastructure, the traditional network perimeter stops being a useful boundary, so the controls have to move closer to the model, the tooling layer, and the execution environment. The first talk, by Nat Natraj (CEO and co-founder of AccuKnox), covers prompt guardrails: what they actually do, the tradeoffs between stateless single-prompt detection and stateful session-aware detection, and the latency and coverage constraints that decide whether a guardrail holds up under production traffic. It includes live demos of bypasses, among them AWS Bedrock guardrails defeated by character injection, indirect prompt injection through a malicious skill in an agentic workflow, and multi-turn jailbreaks such as the Crescendo attack. The takeaway is that ML-based detection is probabilistic, so layering and defense in depth matter more than treating a guardrail as a firewall. The second talk, by Christian Nicholson (lead consultant and partner at Indelible), sketches an architectural blueprint for secure AI infrastructure: Model Context Protocol tooling servers in local versus remote deployments, deterministic policy engines for authorization and tool use, and isolation of untrusted execution using containers and ephemeral microVMs such as Firecracker. A third talk from Roman Trigubenko (technical PM at Wallarm) is still to be confirmed. Doors open at 18:00, talks run from 18:30, and networking fills the final hour before doors close at 21:00.