REA hits 290k GitHub stars, but hands-on tests show AI reverse engineering still falls short
The open-source project REA (Reverse Engineer Anything) has spread rapidly on GitHub, passing 290,000 stars as of October 17, according to MIT Technology Review China. It does not invent a new decompilation technique; instead it connects existing reverse-engineering tools such as Ghidra to coding agents like Claude Code and Codex through MCP, letting the AI call those tools to analyze binaries and attempt to reimplement functionality. When a developer tested it against programs with public source, REA's analysis of the esmtool program from the OpenMW game engine missed real call relationships in 22 of 50 functions while still labeling the output “complete”; in other runs the model skipped REA entirely and gave unverified guesses. REA's author, Morluto, responded in the project's discussions that the prototype-style MCP tool-calling design no longer suits most agents. Separately, the viral ArtCraft apps pitched as open-source alternatives to Adobe and Microsoft products did not use REA; their developer describes the approach as a “clean-room reimplementation” that does not copy the original code.
It shows that wiring professional reverse-engineering tools into AI agents can lower the barrier to this work, while accuracy labeling, code provenance and IP disputes remain practical constraints.
近日,名为 REA(Reverse Engineer Anything)的开源项目在 GitHub 上迅速走红。它声称能够让 AI 在没有源代码的情况下分析软件的内部实现,并尝试重新编写其中的功能。
从办公软件 Notion 的剪贴板机制,到经典游戏 DX-Ball 的声音算法,再到《东方幻想乡》的弹幕计算方式,一系列演示引发了广泛关注。截至 10 月 17 日,GitHub 页面显示,其 Star 数已超过 29 万。科技博主 Matthew Berman 甚至在社交平台上感叹:“Software is done”(软件完了)。
热潮之下,各种 AI 逆向软件的案例也开始传播。有人尝试重建 Claude 的幻灯片编辑器,有人用它给家里的智能马桶开发新的控制工具,还有开发者借助 AI 推出了对标 Adobe、微软多款核心软件的开源应用。
#REA#GitHub#Ghidra#Claude Code#Codex#Morluto#OpenMW#ArtCraft#decomp.dev#IDA#Claude Opus