Tool Pick | reverify — Let Deterministic Byte-Level Tools Be the Judge of AI's Reverse-Engineering Claims

🌏 中文版

Tool Info

What Problem It Solves

Have you ever pointed an AI at an unfamiliar .exe or .dll and asked it to reconstruct a struct, or explain what a chunk of code does? It answers with total confidence — this is the offset, this is what the entry prologue looks like — but you have no way to tell how much of that came from actually reading the bytes versus being recalled from a "textbook answer" buried in training data. Reverse engineering is a worse hallucination surface than reading source code, because there's no compiler standing between a wrong guess and an error message: broken code usually fails to run, but a wrong reverse-engineering claim still sounds completely plausible.

reverify pairs the AI with a pure-Python deterministic reverse-engineering toolkit and makes that toolkit the judge. The model proposes a "claim" — what instructions sit at a given offset, what register values a snippet produces after execution — and that claim only counts once it has actually been checked by a disassembler, a pattern scanner, or a CPU emulator, coming back as VERIFIED, REFUTED, or INCONCLUSIVE along with the evidence bytes. The core is pure Python and installs without Ghidra; running pip install "reverify[full]" upgrades it in place to mature engines — capstone for disassembly, unicorn for real CPU emulation, lief for PE/ELF/Mach-O parsing — and it falls back to the pure-Python core when those aren't installed.

Good fit for: malware analysis, CTF reverse/pwn challenges, interoperability research against a binary you don't have source for — anywhere you want an agent's help reading a binary but can't just take its word for what it says it found.

Quick Start

Install

# Pure-Python core
pip install reverify

# Full version with capstone + unicorn + lief
pip install "reverify[full]"

# Or run straight from a checkout — pure standard library, nothing to install
git clone https://github.com/2akouwu/reverify
python reverify/cli.py auto sample.bin --json

Basic Usage

# Auto-triage: detect format, architecture, sections, extract strings
reverify auto sample.bin --json

# Verify a claim about a specific offset
reverify verify sample.bin --claim '{
  "kind": "instructions",
  "offset": 4096,
  "mnemonics": ["push", "mov", "sub"],
  "note": "function prologue"
}'

The response isn't a bare yes/no — it's VERIFIED, REFUTED, or INCONCLUSIVE along with the bytes the tools actually observed; a REFUTED result also reports where the expected bytes actually turn up. Claims can be batched with --claims-file claims.json, and the CLI exits non-zero if anything is refuted, so an agent or a CI job can gate on it.

Advanced Usage

# Run as an MCP server so Claude Code / Cursor can call it directly
python reverify/mcp_server.py

Over MCP you get two extra tools: re_verify_claim exposes the verification loop directly to the agent, and every grounded result gets written to a per-binary ledger (.reverify/ledger/<hash>.json); re_ledger hands that state back after the host compacts context or runs /clear — because only what the tools actually verified, observed, or refuted was ever trusted in the first place, so nothing is lost by dropping everything else the model said.

Comparison with Existing Tools

Things to Watch Out For

• Authorized use only: the README is explicit that this is for authorized reverse engineering — malware analysis, CTF, interoperability research, or software you own or are permitted to analyze — not a tool for cracking someone else's software.
• Don't over-generalize the benchmark: the published number ("19 real Windows system DLLs, the model's textbook prologue guess wrong 100% of the time, reverify caught every one with zero false accepts") measures one specific and common model prior about entry-point prologues, not a general error rate for all reverse-engineering inference — the project's own BENCHMARK.md says so explicitly.
• Full functionality needs extra dependencies: the pure-Python core's disassembly and emulation coverage is more limited without capstone/unicorn/lief installed.
• Very new project: v0.8.0, created on 2026-08-31 — despite crossing 700+ stars and 148 forks within three days, it's still iterating fast and the interface and claim schema may still change.

Today's Takeaway

Most "AI agent tool verification" stops at whether a tool is allowed to be called. reverify goes one step further and checks whether what the model says about the result of that call is actually true — moving the judgment call from the model's own confidence to the bytes themselves. This pattern isn't limited to reverse engineering: any domain where an agent's output can be checked by a deterministic verifier (type checkers, schema validation, unit tests) can use the same "model proposes, tools decide" loop instead of trusting a model's own claim of certainty.

References

• reverify GitHub repo: project description, README, install instructions, MCP server design, license (MIT), stars/forks — all from the official repo.
• reverify BENCHMARK.md: raw data and methodology for the 19-Windows-DLL prologue-guessing benchmark.
• Model Context Protocol official docs: introduction to the MCP protocol.