ETHPrague 2026

The Premise:
For years, the smart contract security model has been static: Code → Human Audit → Deploy → (Maybe) Hack → Post-mortem. This model assumes that attackers are humans with limited time and attention. This assumption is dead.

We are entering an era where "Offensive AI" agents can decompile bytecode, reconstruct logic, and identify vulnerabilities in unverified contracts faster than a human auditor can read the documentation.

What I will cover:
In this talk, I will share insights from building AI agents that have successfully reproduced the root causes of every major DeFi hack since late 2025.

We will discuss:

1. The Offensive Reality: If our "White Hat" agents can automatically decompile bytecode and generate a PoC for a logic error in minutes, we must assume "Black Hat" agents can do the same. The window for "whitehat rescue" is closing.
2. Beyond Copilots: Why current LLM-based tools (focused on assisting humans) fail against economic logic bugs and complex state dependencies, and why the solution lies in agentic architectures that simulate thousands of attack vectors autonomously.
3. The Data: A breakdown of how agentic systems performed against human experts in recent contests (e.g., Sherlock), identifying where AI is already "superhuman" (deobfuscation, pattern matching) and where it still struggles.
4. The Future Defense: Introducing the concept of "Autonomous Defensive RAG" — a system that doesn't just look for known bugs but continuously learns from the history of all on-chain hacks to predict future attack vectors.

Why this matters:
This is not a talk about replacing auditors to save money. It is a talk about survival. When the attackers are automated, defense must be instantaneous. We will explore what a "Defensive AI" infrastructure looks like for protocols that want to survive the next cycle.