A new survey argues that AI-powered reverse engineering has grown too fast and too scattered to keep straight.
Researchers reviewed 246 papers published since 2015 on using machine learning, large language models, and agentic AI systems to analyze compiled binaries. They sorted the work into 22 distinct reversing domains based on what each system tries to infer, then built a shared taxonomy linking traditional analysis techniques, the artifacts pulled from binaries, how those artifacts get represented for models, and the learning methods applied to them. This is a systematization-of-knowledge paper, not a new tool release, and its stated goal is giving researchers a common vocabulary rather than a new benchmark score to chase.
That matters because binary reversing sits underneath vulnerability discovery, malware investigation, and firmware auditing, and compiling code strips out the human-readable structure that makes those jobs hard in the first place. AI has been recruited to patch that loss piecemeal, with separate subfields inventing their own representations and evaluation practices in isolation. The survey's authors say that fragmentation has made it difficult to compare methods or trust reported results, and they flag evaluation gaps as a recurring problem across most of the 22 domains.
That it took 246 papers and ten years to produce a shared taxonomy is itself a reasonable measure of how improvised AI-security tooling still is.