Large language models first entered reverse engineering as helpful assistants: improving decompiler output, suggesting names, answering questions about code, and generating small analysis scripts. Useful as these capabilities were, they largely remained confined to the role of a smart helper sitting next to the analyst. But reverse engineering has changed quickly. With tool-connected, agentic workflows, models are no longer limited to commenting on pseudocode. They can query disassemblers and decompilers, inspect cross-references and intermediate representations, generate custom scripts, patch binaries, rerun analyses, and iteratively refine hypotheses in a loop. The result is a fundamental shift: from LLM-assisted code reading to agentic reverse engineering.

This shift is especially significant for deobfuscation. Rather than a single clever trick, deobfuscation is usually an iterative process of recovering structure, applying transformations, and reanalyzing the result until something meaningful emerges. Agentic workflows are unusually well matched to this style of work. They can automate common simplification steps, connect different analysis layers, and keep pushing the program toward something more understandable.

Once deobfuscation becomes agentic, it also becomes more global. Instead of operating on one function at a time, agents can help identify core routines, propagate recovered semantics across the call graph, mine repeated obfuscation motifs, and apply transformations at interprocedural or whole-program scale. This changes the practical unit of work in reverse engineering: from manually attacking isolated functions to coordinating larger deobfuscation campaigns across an entire binary or malware family.

In this talk, we revisit the state of LLM-assisted reverse engineering from our earlier work and examine how the landscape has evolved. We argue that the key breakthrough is not that models suddenly understand obfuscated code perfectly, but that they can now help orchestrate the broader workflow around deobfuscation: tool usage, reasoning, scripting, patching, and iterative refinement. This naturally leads to the defensive question as well. If reverse engineering is becoming agentic, what must obfuscation look like to remain resilient against attacks that are tool-driven, scriptable, and scalable? We close by exploring the emerging design space of anti-agentic obfuscation, including intertwined and global obfuscation layers, diversification, and structures that resist clean decomposition and large-scale automation.