At the time of writing, OpenClaw is making headlines every day with new security issues, but also success stories and achievements. There are all kinds of recommendations and suggestions floating around for how to secure your AI agents. Some are confidently explaining that their agent runs on a separate machine that is not exposed to the internet. Others claim that they are safe because their agent has only read access to their emails. Are they really?

In this talk, we break it down and analyze the whole attack surface. We cover everything from prompt injection and memory poisoning to supply chain security and sandbox configurations. The talk maps OpenClaw against OWASP Top 10 for Agentic Applications and provides concrete analysis of each failure mode.

The goal is to learn by example and raise awareness of both new and familiar attack vectors introduced by increasingly autonomous AI agents. These tools have democratized software development in a way we’ve never seen before—which is exciting, but also means we must re‑establish security fundamentals for new audiences, new workflows, and new risks. Attendees will leave with a clearer mental model of where the real dangers lie and how to think about securing agentic systems in practice.