Breaking Obfuscated Binaries with AI Agents: An Attacker’s Playbook

Tim Blazytko explores practical strategies for using tool-using AI agents to deobfuscate binaries and recover hidden logic.

Speaker: Tim Blazytko Part of Binary Cartography, a technical webinar series on reverse engineering, malware analysis, and software protection. Slides, code, samples, and demo environment: github.com/mrphrazer/binary-cartography/tree/main/2026-09-ag… Binary Cartography series repository: github.com/mrphrazer/binary-cartography This webinar explores practical strategies for using tool-using AI agents to attack obfuscated binaries. Rather than focusing on isolated one-shot successes, the session develops workflows for recovering hidden logic, building custom analysis tooling, verifying results, and continuing an investigation when autonomous analysis reaches its limits. The session starts with a bounded control-flow deobfuscation task and develops a workflow around clear analysis objectives and explicit verification. It covers opaque predicates, mixed Boolean-arithmetic (MBA), symbolic execution, and SMT solving, and demonstrates how these techniques can be combined to analyze and simplify protected code rather than merely producing plausible explanations. A second, substantially harder target combines control-flow flattening, arithmetic obfuscation, and wider cross-function dependencies. This illustrates where one-shot analysis stops working, why context and tooling become bottlenecks, and how expert guidance can be used to change strategies, resolve blockers, and make further progress. The webinar also discusses how much autonomy to give different models, how to validate transformations through emulation and other evidence-driven checks, how to preserve analysis state across longer investigations, and how successful techniques can be captured as reusable skills. The goal is not to find a single perfect prompt. It is to develop a practical attacker’s playbook for turning reverse-engineering expertise into repeatable, verifiable, and increasingly scalable deobfuscation workflows. Tools and techniques include Ghidra, Binary Ninja, Miasm, Z3, symbolic execution, SMT solving, emulation, MCP-connected tooling, Docker, Codex, and Claude Code. Designed for reverse engineers, malware analysts, security researchers, and software-protection practitioners comfortable with binaries, disassembly, or decompilation. No prior hands-on experience with agentic analysis workflows is required. #ReverseEngineering #Deobfuscation #AIAgents #BinaryAnalysis #SoftwareProtection 0:00 Introduction to environment 1:02 Tooling and setup 2:15 Defining the deobfuscation challenge 5:22 Strategies and model guidance 8:43 Goal setting and verification 12:26 Automation loops and feedback 16:22 Hands-on agentic deobfuscation 20:18 Project documentation and continuity 22:58 Handling agent failure and intervention 26:38 Advanced analysis: Opaque predicates 33:04 Refining the agentic approach 37:10 Global obfuscation patterns 42:19 Summary and future training