AI-assisted reversing¶

Language model integrations for Ghidra and IDA Pro are useful for speeding up the repetitive parts of binary analysis. They are not a substitute for reading the disassembly.

What works¶

Renaming: feeding a decompiled function to a model and asking for meaningful names for variables and parameters is faster than doing it manually, and the results are often good enough for routine library functions and boilerplate code.

Summarising: a one-paragraph summary of what a function does helps when triaging a large binary with many functions. This allows prioritising which ones to look at more carefully without stepping through all of them.

Pattern recognition: models trained on code can recognise common cryptographic primitives, sorting algorithms, and hash functions from the decompiled output. This is useful when compiler optimisation has made the structure non-obvious.

What does not work¶

They do not handle novel or heavily obfuscated code reliably. A model that has seen thousands of examples of standard AES implementations will produce confident but wrong output when it encounters a custom block cipher or a VM-obfuscated handler. The output looks plausible, which makes it more dangerous than obvious failure.

They have no access to runtime state. A model reasoning about a decompiled function does not know what values the registers or memory held when that function was called. Static decompilation loses that context, and the model cannot recover it.

They hallucinate. Function names, API purposes, and behavioural summaries can be fabricated. Verify anything the model asserts against the actual disassembly before relying on it.

Tools¶

Ghidra: G-3PO and several community scripts pass decompiled C to an OpenAI-compatible API and return annotations. GhidrAssist adds a panel for querying functions directly from the Ghidra UI.

IDA Pro: the Hex-Rays AI assistant (commercial) integrates with the decompiler. Third-party plugins such as Gepetto drive the Hex-Rays decompiler through a language model to provide similar functionality against the OpenAI API.

Binary Ninja: the Sidekick plugin offers model-assisted analysis with a focus on understanding function purpose and data structures.

Workflow¶

ONLY use AI as a first pass on functions not yet examined. Accept the renames and summary as hypotheses, not conclusions. When the function is stepped through or run under a debugger, the model’s interpretation can be confirmed or corrected. Update names accordingly.

The value is in triaging. On a 200-function binary, spending 30 seconds on model-assisted naming for each function and then focusing manual effort on the 10 that look interesting is faster than manually naming all 200.

Do not use model output as evidence in a report. Verify every claim against the binary.