Researchers built a decompiler that checks its own work by recompiling the code it produces.
The system, called Echo, tackles a known weak spot in AI decompilers: their output can look like plausible source code while still being wrong. Echo generates candidate source code and compiler settings with a specialized model, recompiles the candidates, and measures how closely the resulting assembly matches the original binary. When there's a mismatch, it repairs the code through a mix of rule-based rewriting and neural and reasoning-based refinement, then tries again. The researchers tested Echo on function-level benchmarks and on the Mirai botnet malware binary.
On Mirai, Echo matched 2.75 times as many functions as GPT-5.6 and 7.4 times as many as Codex, and it produced 2.43 times more exact matches than the strongest baseline overall. For security researchers who reverse-engineer malware for a living, the difference between code that looks right and code that provably recompiles to the same binary is the difference between a hunch and evidence.
It's a narrow fix, verified matching only works when a working compiler and configuration exists to test against, but it's a rare case in AI-assisted reverse engineering where the tool has to show its work before you trust it.