An offline 27B model broke a paid app's license check in 30 minutes

A model small enough to sit under a desk, with the network cable unplugged, took a paid app’s licensing apart in about half an hour.

What the model actually did

A writer at XDA gave Qwen 3.8 27B a job he assumed needed a frontier model: work out how a commercial application he had legitimately bought verifies its license.

It ran locally on a Lenovo ThinkStation PGX with an Nvidia GB10 chip and 128 GB of unified memory, with no cloud involved at any point. The model performed static analysis of the binary through ARM64 disassembly, recovered the embedded RSA public key, documented the authentication architecture, identified three weak points in the licensing scheme and wrote a working proof-of-concept bypass. It made mistakes along the way and corrected them without being prompted. Total time was roughly 30 minutes.

The caveats the author states

One application, one run, with the reasoning effort set to maximum and heavy token use as a result. He declines to name the app because it is a real product people pay for, and he is explicit that “a harder application might have stopped it completely.” This is a single data point, not a benchmark.

Why the location of the work matters

The interesting part is not the bypass, it is where it happened. Reverse engineering at this level used to mean a frontier API that logs the request and can refuse it. On open weights running offline there is no log, no refusal and no provider in the loop.

That is the same property that makes local models attractive for privacy, which is why the tradeoff is hard to legislate: a model that cannot report you also cannot report anyone else. The open question for open-weight releases is whether capability at this size changes what a license file, and software licensing in general, can reasonably rely on.