Software protection for the age of AI reverse engineering
Language models can now drive a decompiler, read what comes out and explain it. That makes reverse engineering faster and cheaper to attempt. It has not made protection pointless. It has changed which protection is worth paying for.
Updated
What did AI change about reverse engineering?
AI changed the throughput of reverse engineering, not its capability. A coding agent with a shell can run a decompiler, read thousands of functions, write a script to decode what it finds and try again, all without a person who knows assembly. The methods are the ones experts already used. What is new is that anyone can run them, many times over, for the price of a subscription.
Public research this year shows the pattern clearly. In Quarkslab’s tests, an agent handed progressively hardened binaries took the same route every time: find the protected data, find the code that decodes it, copy that code into a script, run it. It never actually removed a protection. It looked for the cheapest path to something that resembled an answer, and took it.
That is good news and bad news. The bad news is volume: a vendor should expect automated, repeated probing of every release, not one patient attacker. The good news is that an agent looking for shortcuts is defeated by protection that offers none.
Does obfuscation still work against AI?
Yes, as a cost multiplier. It never worked as a wall, and it still does not. Elastic Security Labs ran a frontier model against a small program compiled through 22 combinations of an academic obfuscator. It solved the unprotected version every time, about half of the single-transform versions, and none of the heaviest three-layer stacks. Where it did succeed against hardened builds, it took up to about four times the cost and time.
The lesson is not that one technique is unbreakable. It is that stacked, varied transforms make reading so expensive that the agent has to fall back on running the code. That moves the fight to runtime, where the program can defend itself.
Simple renaming is the exception. Scrambled identifiers slowed people down, but a model recovers intent from structure and calls. Renaming alone now buys almost nothing.
What does real software protection look like now?
Real protection is three layers that depend on each other: obfuscation, runtime anti-tamper, and licensing the server decides. Each one covers the gap the others leave.
- Layered obfuscation makes static reading expensive enough that an attacker has to execute the program to learn anything. Read what AI agents can and cannot get through.
- Runtime anti-tamper, also called runtime application self-protection, notices debuggers, instrumentation and modified code while the program runs. It should degrade quietly rather than crash, because a crash tells the attacker exactly which check fired.
- Server-side licensing keeps the decision about who may run the software on a server you control, so there is no local check to patch out. Read why the license check belongs on the server.
The first two raise the cost of an attack. Only the third removes the prize: if the build does not contain what unlocks it, recovering the build is not enough.
Why does protection need to be part of the distribution platform?
Because a protected build is only as strong as the licensing and delivery around it. A heavily obfuscated client behind a license key that is checked locally is still one patch away from free. A build that anyone can download can be studied at leisure.
So the parts have to share one source of truth. Licenses are verified by the server. Builds are signed and delivered only to accounts entitled to them. Each copy carries a watermark tied to the license it came from, so a leak has a name on it. Telemetry from the field shows tampering while it is happening, which matters when attempts are automated and constant.
That is how the Cerberus platform is built: protection, licensing, distribution and telemetry run as one system for each tenant.
How does Cerberus approach it?
We design a protection pipeline for each product and aim it at automated analysis first. We test protected builds against current models and harden against what they recover. Builds vary from release to release, so whatever an attacker learns about one does not carry over cheaply to the next.
Specifics: .
We do not publish the methods. A published method becomes training data, and a model that has read it knows where to look.
What can't protection do?
No protection on a user’s machine keeps a secret forever. Someone patient enough, with enough time, can read anything that runs where they control the hardware. Protection is worth paying for when it makes that cost higher than the software is worth, and when it is paired with secrets that never leave the server.
Some of today’s advantages will also narrow. Agents are weakest when they must execute code rather than read it, and their execution tooling is improving. Protection has to be maintained like any other security control, not applied once.
Questions
- What is AI-assisted reverse engineering?
- It is reverse engineering where a language model does the reading. An agent with shell access runs a disassembler or decompiler, reads the output, writes scripts to decode or emulate what it finds, and reports back. It works from the compiled program, not the source.
- Did AI make obfuscation useless?
- No. Public 2026 studies found that agents route around obfuscation rather than undoing it, and that layered obfuscation raised their cost several times over or exhausted their budget entirely. Obfuscation still works as a cost multiplier. It was never a wall, and it is still not one.
- What changed, then?
- Throughput. An agent can make attempts in parallel, overnight, without an expert at the keyboard. The techniques are the ones a skilled reverse engineer already had. Defenders should assume many more attempts by less skilled people, not a new kind of attack.
- What does good protection look like now?
- Three layers that depend on each other: obfuscation to make static reading expensive, runtime anti-tamper so the program notices analysis and degrades without revealing which check fired, and licensing the server decides, so the value is not inside the build at all.
- Can any client-side protection keep a secret forever?
- No. Anything that ships to a machine you do not control can eventually be read by someone patient enough. The goal is to make that cost more than the software is worth, and to keep the secrets that matter on a server.
Sources
- Defeating AI-Assisted Reverse Engineering (or at Least Trying To), Quarkslab, August 2026
- The Cost of Understanding: LLM-Driven Reverse Engineering vs Iterative LLM Obfuscation, Elastic Security Labs, April 2026