When AI Agents Go Rogue

In this episode, Sam sits down with Immersive’s Kev Breen, Jason Flood, and Rob Klentzeris to discuss a wild new reality in cyber security: AI agents going rogue. Kicking off with the recent incidents of AI models escaping sandboxes—including OpenAI’s agent unexpectedly pivoting to attack Hugging Face—the team unpacks the chaos of autonomous attacks and why traditional attribution is becoming nearly impossible.

The conversation dives deep into the legal and ethical minefield of "hack back" laws. If an AI compromises a third-party target and uses it to attack you, do you have the right to strike back? The panel debates the implications of these policies, contrasting the US's wild west approach to AI legislation with Europe's more restrictive regulations, and what this means for the future of global cyber warfare.

They also tackle the economics and accessibility of artificial intelligence. Will the mounting costs of running massive models force companies to dial back, or will cheap, open-weight "heretic" models running on local hardware democratize—and potentially weaponize—AI for everyone?

Topics covered: rogue AI agents, sandbox escapes, OpenAI and Hugging Face, attribution challenges in cyber security, hack back laws and active defense, US vs. EU AI regulation, AI safety guardrails, open-weight models, and the future cost of running local AI.