AI Reverse Engineering Gains Momentum: From Creative Apps to Classic Games
AI-assisted reverse engineering is attracting attention because agents can now take part in the full investigation loop: inspect a target, propose a behavioral rule, implement it, and compare the result. [REA](https://github.com/morluto/rea), the project currently featured in sAIgnals' [repository view](/repositories), had more than **122,000 GitHub stars when checked on 12 October 2026**. That is evidence of interest in the tooling; it does not measure reconstruction quality or production adoption.
The practical story is visible in three different projects. Two are independent creative applications that reproduce familiar workflows across platforms. The third documents a reconstruction from an original game executable with REA. Keeping those distinctions clear makes the trend more useful than the slogan “AI can clone anything.”
PhotoCraft brings a familiar image-editing workflow to open source
[PhotoCraft](https://github.com/storytold/photocraft) describes itself as a clean-room Photoshop reimplementation written in Rust. It offers familiar layers, masks and document workflows, with builds for several desktop platforms and the web. Its developers say they work from public specifications and observed behavior, and its PSD support is tested against reference files.
This is an independent application, rather than Adobe's source code ported to another operating system. The project's README still calls it **early alpha** and describes substantial gaps in professional workflows, tools and compatibility. Familiar menus and an impressive demo do not establish that an existing production workload will transfer unchanged.
VectorCraft tackles the Illustrator workflow
[VectorCraft](https://github.com/storytold/vectorcraft) applies a similar approach to vector illustration. The Rust project targets desktop platforms and the web, with drawing, paths, paint and document interchange. Its public roadmap and source make the reconstruction inspectable.
The project is also **alpha**. The maintainers identify missing or incomplete areas, and their feature-parity estimates are self-assessments. Treat those estimates as a development map, then test the documents and editing operations that matter to you. Neither PhotoCraft nor VectorCraft's cited README establishes that REA was used to build it; they illustrate the broader reimplementation trend.
DX-Ball shows what a checked REA reconstruction looks like
The [DX-Ball 1.07 reconstruction](https://github.com/N0zoM1z0/dx-ball) explicitly documents using Codex with REA to inspect the original Windows executable and reconstruct readable C/C++. Its workflow compares recovered code against original behavior and selected compiled-byte oracles.
REA's [sound-panning case study](https://rea.tools/showcase/dx-ball/) gives a concrete result: a recovered helper is reported to pass **3,205 original-x86 cases** and reproduce **all 63 compiled bytes** of the function. This is a project-reported check for one function, not an independent benchmark or a claim that every part of the game is complete. Still, it shows the evidence an agent-assisted reconstruction can provide beyond a video of something that looks similar.
What has changed for developers?
REA connects an agent to analysis tools through MCP, with a CLI for direct inspection. Native targets can use established providers such as Ghidra, Hopper or IDA; static JavaScript/Electron inspection has a simpler path. The agent can use the returned findings to ask follow-up questions and write a replacement. The engineering task remains defining which behavior must match and checking it against the target.
These examples suggest a useful direction for legacy maintenance, interoperability and software preservation. A recovered parser, file format, calculation or IPC contract can be more valuable than a sweeping application clone. The result to look for is a behavioral specification and a reproducible comparison, with unknowns recorded.
The project's own license does not grant rights to redistribute another application's code or assets. REA's [responsible-use guidance](https://github.com/morluto/rea/blob/main/DISCLAIMER.md) frames it around permitted research, education, interoperability and preservation. Start with a target you can legitimately inspect and a bounded feature you can verify.
For a concrete first run, our [REA guide for Claude Code and Codex](/articles/how-to-use-rea-reverse-engineering-claude-code-codex) connects the tool, traces an Electron CSV export, and checks an input containing a comma and quotation marks. It turns the trend into a small development workflow you can reproduce.