AI agents are under attack. From prompt injection exploits to invisible system takeovers, new security threats are forcing a rethink of how we build, test, and trust autonomous systems.

Welcome to Episode 23 of the Agents Unleashed Podcast, the show that helps you find signal in the noisy world of agentic AI.

Hosted by Thomas Maybrier, this episode investigates the growing danger of prompt injection, and how attackers are learning to hijack AI agents to steal data, drain wallets, and impersonate users. But it’s not all bad news: Thomas also explores how open-source workflows, decentralized protocols, and new evaluation tools like Olas Predict may offer a more secure path forward.

In This Episode:

Why prompt injection is the #1 threat for AI agents

Real-world hacks from BlackHat

What red teaming has revealed about agent behavior

How Olas agents handle trust-minimization, verification, and incentives

Chapters

00:00 – Welcome to Agents Unleashed

00:50 – Real-world prompt injection at Black Hat

03:21 – NVIDIAGTC demo: multi-agent red teaming

04:07 – Why prompt injection is a systemic threat

04:56 – The risks of compromised agents

05:16 – How do we make AI agents trustworthy?