Global AI Karlsruhe

🤖 Agent Smith: From Exploitation to Containment

AI agents are no longer just generating text — they act.
In this two-part session, we first let autonomous agents loose on a real IoT device, then try to bring them back under control in an enterprise environment.

🛠 Part 1: Agent Smith Gets Hardware:
Autonomous IoT Hacking From Debug Port to Cloud API

Speaker: Marc Plogas

Last year, Agent Smith turned rogue in software. This year, he got physical access. In this session, we take a cheap off-the-shelf IoT device and let AI agents loose on it. No manual hacking, no memorized commands, no Kali cheat sheets. Using MCP-driven tool orchestration, the agents autonomously probe the device's hardware debug interface, extract firmware secrets, and intercept its cloud traffic through a rogue WiFi access point. Multiple attack vectors, one device, zero human intervention.

This is not a theoretical exercise. Every demo runs on a Raspberry Pi with open-source tools, and the agents make their own decisions about what to probe, what to extract, and what to flag. We will see what they finds, what they miss, and what happens when a device actually fights back.
Whether you build IoT products, secure enterprise networks, or just want to know what your smart devices are doing behind your back: this session will make you uncomfortable.

🛡 Part 2: Agent Smith Gets Governed
Securing AI Agents with Microsoft Foundry, Control Planes & Guardrails

Speaker: Dr. Alexander Wachtel

After watching Agent Smith break things, we face the harder problem: how do we control it? Modern AI agents don’t just generate text—they act. They call APIs, orchestrate tools, access data, and make decisions. That makes them powerful—and dangerous. In this session, we explore how to secure autonomous agents using Microsoft Foundry, the Foundry Control Plane, and built-in guardrails. We treat agents as active actors in your infrastructure, not just interfaces.

Through real scenarios, we demonstrate how to:
- Enforce identity and access boundaries
- Restrict tool usage and execution paths
- Prevent data exfiltration
- Monitor and intervene in real time

We’ll also look at where guardrails fail—because autonomous systems don’t always break in predictable ways.

⚡ What to Expect
- Live demos (real hardware & real agents)
- Autonomous attacks and defensive controls
- Practical insights into AI agent security
- Honest results: what works, what doesn’t

👥 Who Should Attend
- Security professionals & red/blue teams
- Architects & IT leaders
- Developers & AI engineers
- IoT / embedded engineers