Reverse engineer in Binary Ninja — HLIL/MLIL/LLIL inspection, strings/imports/exports, cross-references, types, patching, Python API automation, and the optional MCP/HTTP bridge. Load when the user picks Binary Ninja, when its IL levels materially help data-flow analysis, or when IDA/Ghidra/radare2
v2026.10.14-ee0e0c53 установки+3 за 7 дней★ 27 ЗвездыMITсинхронизировано 22 часа назад
You're analyzing an in-scope binary (scope.txt / client-owned / your own sample —
tradecraft-scope-roe) and Binary Ninja is the chosen tool, its intermediate languages genuinely help
(SSA-form data-flow questions), or a high-impact conclusion from reverse-eng-ida,
reverse-eng-ghidra, or reverse-eng-radare2 deserves a second decompiler's opinion. Binary Ninja is
commercial software — it needs a valid Vector 35 license, installed by the user.
Binary Ninja lifts every function into layered ILs — LLIL close to the machine, MLIL in SSA form
(variables versioned, so data flow is explicit), HLIL close to source. Questions like "where does this
attacker-controlled value come from" become IL queries instead of manual register tracking, and the
whole analysis is scriptable through one Python BinaryView API.
Baseline the target first — record file hash, architecture, entry points, segments,
imports/exports, and representative strings before touching anything. Work on a copy whenever you
intend to patch or save database changes.
Pick the integration:
GUI / Python API — preferred when Binary Ninja is already open or analysis is interactive.
MCP bridge — only when explicitly requested. The reviewed community option is
fosdickio/binary_ninja_mcp (GPL-3.0, not an official Vector 35 component; min Binary Ninja
build 4000). Keep its HTTP listener on 127.0.0.1:9009 — never expose it to the network.
Pin the bridge version:text
The bridge only works once Binary Ninja is running with a binary open and the plugin endpoint
responds; register it only in the MCP client the user chose. Discover the live tool list rather
than assuming every upstream function exists (families include view selection, list_imports,
list_exports, list_strings, decompile_function, get_il, callers/callees, xrefs, types,
comments, renames, byte patching).
Navigate by references, not linearly. Follow call sites and cross-references before
interpreting any function in isolation.
Pick the right IL level per question: HLIL for readable logic, MLIL SSA for data-flow tracing,
LLIL/disassembly when the lifting loses instruction-level behavior (self-modifying code, exotic
instructions, anti-analysis tricks).
Annotate incrementally — names, comments, types — keeping original addresses in your notes so
every renamed symbol stays traceable.
Treat mutations as mutations. Byte patches, prototype changes, and saved-file writes happen
only when requested, on a preserved copy of the original artifact.
Cross-check high-impact conclusions with a second evidence source or another disassembler
before they go in a report.
Report concrete addresses, function names, the IL level each conclusion was drawn at, supporting
strings/imports, a confidence note, and reproduction steps — so a reviewer can re-walk the analysis in
the GUI and land on the same functions.