A lie detector for AI reverse engineers

What it does

Reverify is an AI-assisted reverse engineering toolkit that uses deterministic analysis as the final authority. A language model can hypothesize about structures, algorithms, or offsets inside a binary, but the claim only survives if the toolkit verifies it against actual bytes through disassembly, pattern matching, or CPU emulation. It ships as both a CLI and an MCP server, so agents like Claude Code can propose claims and have them judged by the binary itself.

The interesting bit

The verification loop is the heart of it: every hypothesis returns VERIFIED, REFUTED, or INCONCLUSIVE alongside the observed evidence, and the reconstruct command closes the loop by letting a model iterate until its output is grounded. The toolkit is pure Python out of the box—handling PE/ELF/Mach-O parsing, disassembly, and emulation—but upgrades itself in place to Capstone, Unicorn, and LIEF when installed, falling back to the hand-rolled core when they are absent.

Key highlights

• Deterministic judge: claims about instructions, emulate_result, bytes_at, and more are checked against the binary, not the model’s confidence.
• Pure-Python core with optional heavyweight engines: runs without Ghidra or external dependencies, but upgrades to Capstone, Unicorn, and LIEF when available.
• Agent-native via MCP: exposes the toolkit and verification loop to AI agents, so hypotheses can be checked before they reach the user.
• Closed-loop reconstruction: the reconstruct command iterates model proposals through verification until the output is grounded.
• Broad binary support: PE/ELF/Mach-O parsing, x86/x64/ARM/ARM64 disassembly, AOB scanning, Protobuf/TLV dissection, and Frida hook generation.

Caveats

• The project is explicitly scoped to authorized reverse engineering only; the authors flag malware analysis, CTFs, and interoperability research as acceptable use cases.
• It is early days—v0.3.0 with 96 unit tests—and the README does not discuss performance or scalability limits for large binaries.

Verdict

Worth a look if you are using AI agents for binary analysis and need a way to stop them from inventing offsets. Skip it if you want a mature GUI disassembler; this is a verification layer, not a replacement for IDA or Ghidra.