研究人员发布了Kuna实验性反编译器，该项目在美国空军研究实验室（AFRL）和Metalware的支持下由大语言模型开发[1]。Kuna在C程序控制流结构化任务上的完美结构率达到44.4%，接近业界成熟工具IDA Pro的45.7%[1]。项目代码几乎完全由LLM编写，通过自主学习和迭代优化实现了包括20多项angr基础功能的重新实现[1]。
Kuna以NSA开源工具Ghidra的Rust版本为基础进行改进，结合了二进制分析框架angr的管道架构设计[1]。LLM通过识别Kuna相对于IDA Pro在基础指标上的不足，自主进行了优化[1]。
Researchers have released Kuna, an experimental decompiler developed primarily by large language models with support from the Air Force Research Laboratory (AFRL) and Metalware.[1] The project demonstrates the capability of LLMs to autonomously learn, iterate, and optimize software development tasks.
Kuna achieves a perfect structure rate of 44.4% when analyzing control flow structures in C programs, approaching the performance of the industry-standard IDA Pro at 45.7%.[1] The decompiler represents a Rust-based improvement over NSA's Ghidra, incorporating the pipeline architecture of the angr binary analysis framework.[1] Nearly all of the project's code was written by the LLM, which successfully reimplemented over 20 foundational angr capabilities.[1] Through self-directed learning, the LLM identified and addressed performance gaps relative to IDA Pro, continuously refining Kuna's output across development iterations.[1]