# Zion Basque releases Kuna, a decompiler built almost entirely by an LLM > Source: > Published: 2026-07-30 04:01:10+00:00 [Zion Leonahenahe Basque](https://noelo.org/people/?ref=runtimewire) released [Kuna](https://kuna.noelo.org/?ref=runtimewire) on July 29th, opening the code for an experimental decompiler that he says was written almost entirely by a large language model. The project is Basque's test of a specific claim about coding agents: a researcher can direct them with measurements and examples of failure, then let the agents handle most of the implementation. In [his release post](https://noelo.org/blog/kuna-release/?ref=runtimewire), Basque wrote that an LLM "has written nearly every line of code" in Kuna. His contribution sits above that code generation loop: choosing the architecture, developing the metrics, assembling the benchmark data and deciding which shortcomings the agents should attack. That division of labor reflects Basque's path into the project. He completed a Ph.D. at Arizona State University with a dissertation titled ["Toward a Science of Software Reverse Engineering"](https://www.zionbasque.com/files/dissertation.pdf?ref=runtimewire), arguing that reverse engineering needs clearer goals and measurable definitions of success. He spent his doctoral work as a core developer of the angr decompiler and studied how metrics can guide decompiler design. [Noelo Lab](https://noelo.org/people/?ref=runtimewire), which Basque will direct at the University of Georgia, formally starts on August 1st. Kuna turns that research agenda into a working tool. Basque built it over the summer while serving as a visiting faculty researcher at the [Air Force Research Laboratory](https://www.afrl.af.mil/?ref=runtimewire) and a research fellow at [Metalware](https://www.metalware.com/?ref=runtimewire), a San Francisco firmware-security developer from [Y Combinator's Summer 2023 batch](https://www.ycombinator.com/companies/metalware?ref=runtimewire). Kuna remains a Noelo Lab research project rather than a Metalware product. ### A decompiler designed for its own coding agents Decompilers convert machine code into higher-level representations that humans and software agents can inspect. Their output underpins malware analysis, vulnerability research and work on programs whose source code is unavailable. [Kuna's source code](https://github.com/Noelo-Lab/kuna?ref=runtimewire) began as a Rust port of the National Security Agency's [Ghidra](https://github.com/nationalsecurityagency/ghidra?ref=runtimewire) decompiler. Basque then reshaped the pipeline around ideas from angr and other decompilation research. The Apache-2.0 project runs from a command line, inside a browser through WebAssembly and as a decompiler core within Ghidra's graphical interface. Its architecture is meant to make further changes easier for agents. Pipeline phases are separated, important algorithms receive natural-language specifications, and decompiler features can be toggled at runtime. That gives an agent a smaller search space when a benchmark exposes a bad output. It can locate the relevant phase, inspect the documented behavior, modify a feature and run the tests again. Basque says that process allowed agents to reimplement [more than 20 features](https://github.com/Noelo-Lab/kuna/tree/main/docs/features?ref=runtimewire) previously developed for angr. The historical dependency matters. Kuna draws on techniques produced through years of human-led research, including Basque's own work, and packages them in a codebase structured for repeated agent modification. The project repository makes the intended operating model explicit. Coding agents are expected to perform most code analysis and creation, while humans supply design judgment, datasets, prompts and acceptance criteria. Kuna's priorities also differ from conventional reverse-engineering tools: decompiled text quality and speed rank above graphical interfaces because Basque expects software agents to become major consumers of decompiler output. ### The benchmark result comes with a controlled denominator Basque's headline result is close enough to demand attention and narrow enough to require care. On C programs in [DecBench](https://decbench.com/?ref=runtimewire), Kuna produced perfect control-flow structuring for 44.4% of functions, compared with 45.7% for IDA Pro 9.2, according to Basque's [release post](https://noelo.org/blog/kuna-release/?ref=runtimewire). Control-flow structuring determines how low-level jumps and branches are reconstructed into readable constructs such as loops, conditionals and switches. It is one component of decompilation quality. Kuna's result does not establish parity with IDA Pro as a complete reverse-engineering system, and Basque does not present it that way. He lists types, optimization, recompilability and variable identification among the areas that still need substantial work. The benchmark also comes from Basque's own research environment. DecBench identifies itself as an experimental, living benchmark and acknowledges that bugs or overrepresented projects may affect its results. As of July 30, 2026, [DecBench lists](https://decbench.com/?ref=runtimewire) 94,575 functions across 803 binaries and 9 decompilers, scored on three metrics: Code Structure, Types (args and vars), and Byte-match Recompilability. Independent reproduction will determine how well Kuna's structuring result holds beyond that setup. Still, the comparison demonstrates something useful about agent-led engineering. The agent was given concrete examples where Kuna underperformed IDA Pro, Ghidra or angr, then asked to refine the implementation through trial and error. The feedback came from measurable output rather than a general instruction to improve the decompiler. That loop places the difficult human work at the beginning. Researchers must decide what quality means, build representative datasets and identify metrics that correspond with the needs of reverse engineers. Basque's dissertation and related research supplied those ingredients before an agent could start making useful changes. ### Basque is keeping the research loop open Basque is also keeping angr at the center of his work. He described it as his preferred environment for developing new decompilation algorithms, while Kuna tests how agents can absorb and refine established techniques. The relationship is deliberately reciprocal. Kuna depends on angr and other open-source research today; Basque's goal is for improvements produced through Kuna eventually to feed back into that research base. Metalware provides an adjacent commercial setting for the work. Metalware analyzes firmware binaries without requiring source code, hardware or test harnesses, then returns crashes, stack traces, reproducible inputs and root-cause analysis. Better automated decompilation can give security agents clearer representations of the binaries they are testing, even though Kuna is separate from Metalware's product. Kuna's immediate value lies in the experiment Basque has made inspectable. The code, specifications and test structure show where he believes coding agents can carry implementation work and where a domain expert remains essential. The early benchmark result gives that argument a measurable checkpoint. The harder tests are still ahead: whether agents can improve the less mature parts of the decompiler, whether outside researchers can reproduce the gains and whether Kuna can produce techniques worth sending back upstream.