AI Reverse Engineering: What REA and Ghidra Can Do

Two weeks ago, reverse engineering was a niche skill that took years to learn. Then an open-source project called REA, short for Reverse Engineer Anything, went from launch on 7 October to the top of GitHub, and timelines filled with claims that AI could now take apart any app, game or console. Some of that is true. This guide covers what AI reverse engineering actually is, what REA does and does not do, how good the models are, and where the legal lines sit, including in India.

What AI reverse engineering means

Reverse engineering is working out how software behaves when you do not have its source code: reading the compiled machine instructions, naming functions, following data, and testing guesses. The classic tools are disassemblers and decompilers such as Ghidra, IDA and Hopper. They turn a binary into something a person can read, but a person still has to read it.

The change in 2026 is that an AI coding agent can now do the reading. Given access to those tools, an agent can open a binary, inspect functions and cross-references, form a hypothesis about what a routine does, write code that reproduces it, and run tests to check. The human sets the goal and judges the result.

What REA is and how AI reverse engineering works with it

REA is a free, MIT-licensed command-line tool and MCP server made by a developer who goes by morluto. MCP, the Model Context Protocol, is the standard way AI agents call outside tools. According to its README, REA connects agents including Claude Code, Codex, Cursor, Gemini CLI and Grok Build to three native analysis back ends, Hopper, Ghidra and IDA, and also inspects JavaScript and Electron apps, .NET assemblies, Android APKs, firmware, packet captures and SQLite databases. Each conclusion comes with an evidence record, so you can see what the agent looked at.

The numbers explain the noise. GitHub's own data showed 328,745 stars and 83,218 forks on the morning of 19 October 2026. Startup Fortune reported about 10,400 stars a day after the 7 October launch.

Two practical notes from the documentation and from reviewers:

• It runs locally, but your model provider still sees what you send. REA does not upload your target, yet the agent you connect sends its prompts and findings to that agent's model provider under its own data policy.
• Pin the version. A security review on skills.sh rated REA's skill medium risk because a setup path runs the latest package without a fixed version and downloads a disassembler. Install a specific version and read what it fetches.

What it can and cannot do

The showcases in REA's documentation, also described by InfoQ, are specific and checkable:

• DX-Ball: the agent reconstructed a sound-panning function from the classic game, with tests matching all 3,205 original x86 cases and all 63 compiled function bytes.
• Notion: it traced how the Electron desktop app moves data through the clipboard.
• Touhou 4: it recovered the bullet-ring angle calculation from 16-bit DOS code.

What it does not do matters as much. InfoQ notes that REA "does not claim to recover the original source code", and the README says it does not clone applications automatically. It explains behaviour and can rebuild individual functions that pass the same tests. That is useful, but it is not the same as getting a program's source.

The viral claim that AI had cracked PlayStation 5 security "in weeks" is a good test case. Towards AI's analysis points out that a separate kernel exploit had already been published in late September, before the screenshot spread, and that the AI's share of the result was never shown. Treat such claims as unproven until someone publishes the evidence.

How good are the models at it?

OpenAI's GPT-6 Astra announcement reports results on SRE-Bench, a test of whether a model can reverse engineer software binaries:

Model	Solved on one attempt	Solved within four attempts
GPT-6 Astra	88.0%	99.2%
GPT-5.6 Sol	55.9%	68.7%
Claude Opus 5	12.5%, attempts not stated	—

Two cautions. These are OpenAI's own numbers on its own setup, and the newsletter Popular AI notes the strongest results used an unusually large compute budget across repeated attempts, so they do not mean "Astra can reverse engineer 99% of software." OpenAI also says Astra refuses more advanced cyber tasks such as writing proof-of-concept exploits.

The AI reverse engineering toolkit at a glance

Tool	What it is	Cost
REA	Bridge that lets AI agents drive analysis tools	Free, MIT licence
Ghidra	NSA-built decompiler and analysis suite for Windows, macOS and Linux	Free, Apache 2.0
IDA Pro	Long-standing commercial disassembler and decompiler	Paid
Hopper	Disassembler REA uses for deep native analysis	Paid
Claude, ChatGPT via Codex, Cursor	The agents that do the reasoning	Their usual plans

Ghidra is the sensible starting point: the NSA's Research Directorate maintains it, it is open source, and it is the back end most free guides assume. For which agent to pair it with, our ranking of AI coding tools compares the options.

Is AI reverse engineering legal?

It depends on what you analyse, why, and the licence you agreed to; this is not legal advice. In India, Section 52(1) of the Copyright Act, 1957 lists acts that do not infringe copyright in a computer programme. Clause (ab) covers acts by a lawful possessor that are necessary to obtain information essential for interoperability with an independently created programme, where that information is not otherwise readily available. Clause (ac) covers observing, studying or testing how a programme functions to determine the ideas and principles behind it.

Those exceptions do not cover copying a paid product, unlocking paid features or distributing what you extract, and a software licence can add its own limits. REA's README draws the same line: it is for "lawful research, education, interoperability, software preservation" and for software you own or are authorised to examine.

Who it is for, and what software makers should do

For defenders the case is strong. India's CERT-In warned on 15 October that criminals are already using AI attack kits that cut the time from break-in to damage to hours, as our report on the CERT-In advisory explains. Malware analysts and incident responders need the same speed; our guide to malware reverse engineering with AI covers the benchmark results and a safe lab setup. Teams maintaining old binaries whose source was lost, and developers building tools that must interoperate with closed formats, also gain.

For software makers the lesson is to stop relying on compiled code being hard to read. Assume anything shipped to a user's machine can be understood. Keep secrets, API keys and licence checks on the server, not in the app, which is the same exposure CERT-In found AI agents exploiting.

FAQ

What is AI reverse engineering?

Using an AI agent to drive reverse-engineering tools such as Ghidra or IDA, so the agent inspects compiled software, explains what functions do and rebuilds them with tests. A person still sets the goal and checks the result.

Is REA free?

Yes. REA is open source under the MIT licence. The AI agent you connect to it, such as Claude Code or Codex, needs its own plan, and some back ends, such as IDA Pro and Hopper, are paid.

Can AI decompile an exe back to source code?

Not to the original source. Tools can produce readable decompiled code and agents can rebuild individual functions that behave the same, but REA's own documentation says it does not recover original source code.

Is reverse engineering legal in India?

Section 52(1)(ab) and (ac) of the Copyright Act, 1957 allow a lawful possessor to study how a program works and to obtain information needed for interoperability. Copying paid software, bypassing paid features or distributing extracted code is not covered, and licence terms may restrict more.

Which AI is best for reverse engineering?

On OpenAI's SRE-Bench figures, GPT-6 Astra leads, at 88.0% on one attempt. Those are vendor-reported numbers on a single benchmark, so test the agent you already use on your own binaries before switching.