AI models demonstrate advanced reverse-engineering capabilities

Recent experiments highlight the evolving capabilities and limitations of large language models in reverse-engineering and cybersecurity tasks.

Researchers at Quarkslab conducted tests using Claude Opus 5 to analyze hardened AArch64 binaries. The study found that rather than successfully deobfuscating complex code paths, the AI agent tended to route around protections. It achieved its goals by using dynamic analysis tools like QEMU and Unicorn to execute code snippets or by searching the workspace for clues, rather than fully untangling the obfuscated logic. This suggests that while AI can assist in analysis, it may prioritize accessible workarounds over verified deobfuscation.

In a separate demonstration, the Qwen3.8-27B open-weights model successfully reverse-engineered a commercial application’s license check. The model recovered an obscured cryptographic key from ARM64 assembly and produced a working proof-of-concept bypass within approximately thirty minutes on desktop hardware. This capability underscores the potential for smaller, high-performance models to perform complex, agentic coding and security tasks.