Cybersecurity in the Golden Era of AI

Abstract

On May 16, 2026 I served as a panelist at the AI & Cybersecurity Forum in Bogotá. The panel — four engineers plus a moderator — covered "Cybersecurity in the Golden Era of AI," focused on agent models, emerging threats, and prompt injection attacks.

The conversation traced how traditional security frameworks struggle to capture the unique risks of LLM-powered systems, and where engineering teams should concentrate mitigation effort as agent autonomy increases.

This page is the canonical record of the appearance. The full recording, slides, and a written recap will be added as they come together.

Recording

Key Takeaways

• Traditional threat frameworks weren't designed for LLM autonomy. STRIDE and similar models treat the AI as a tool, not an actor. Agentic systems need threat models where the model itself is an attack surface with its own trust boundary.
• Prompt injection is the SQL injection of the AI era — it is not a model problem, it is an input validation problem. External content (emails, web pages, user messages) must be treated as untrusted input even when processed by an AI agent.
• The most dangerous agentic deployments are those where humans trust the output without reading it. The security perimeter follows the trust boundary, not the system boundary.
• Most AI security failures in production aren't exotic. They're agentic systems with over-broad tool permissions executing instructions injected from user-controlled content — a well-understood category of vulnerability running on new infrastructure.
• Defense-in-depth for AI agents means minimal tool permissions, output validation before irreversible actions, human-in-the-loop for high-stakes operations, and adversarial red-teaming before production deployment.

Photos