We demonstrated live how an AI agent can be made to push a simulated control center into a critical state.
At the heise devSec conference in Marburg, we demonstrated live how an AI agent can be made to push a simulated control center into a critical state until the warning light comes on. In this episode, we explain how the test setup is configured and why the attack works without any technical exploit. Because as soon as agents are allowed not only to provide information but also to take action, the security question is no longer what data a system reveals, but what an attacker can make it do.
We're talking about real-life incidents, the limitations of traditional access control concepts, and which control patterns are effective when a model is part of the control path.