Kuna: Decompiler Development in the Age of Coding Agents

Kuna is an experimental "agent-first decompiler designed for autonomous refinement," announced by Zion Leonahenahe Basque (UGA / AFRL / Metalware). Its central claim: an LLM wrote nearly every line of Kuna, yet it rivals IDA Pro 9.2 in control-flow structuring on C programs.

On decbench.com benchmarks, Kuna achieves perfect structuring on 44.4% of functions versus IDA's 45.7%. This was reached through autonomous refinement — the LLM studies cases where it underperforms IDA and iteratively improves. The same process reimplemented more than 20 fundamental features from the angr decompiler that took years of human research.

Basque names three limitations. First, Kuna is possible only because of decades of prior decompiler research; technically it's a Rust port of Ghidra reworked to match angr's pipeline, and angr remains the home for frontier algorithms. Second, autonomous refinement still depends on human scientific insight — choosing metrics, benchmarks, and what constitutes meaningful improvement. Third, the work so far covers only control-flow structuring; types, optimizations, recompilability, and variable identification remain open questions for the framework.