Researchers develop multi-stage AI model to decode JavaScript obfuscation more reliably

Researchers have proposed a novel approach to reversing JavaScript obfuscation—a common technique that scrambles code to prevent analysis—by having AI models learn from the multi-step process that human deobfuscation tools use.

The Challenge of Obfuscated Code#

JavaScript obfuscation protects intellectual property but creates substantial obstacles for security analysis, code review, and program understanding. Existing machine learning approaches to deobfuscation treat the task as a single translation step, overlooking how practical deobfuscation actually unfolds in stages.

A New Training Strategy#

The research introduces a stage-aware latent reasoning framework that captures intermediate outputs from deterministic deobfuscation tools and converts them into training signals. The model learns to follow these multi-stage rewrites during training, yet produces fully deobfuscated output in one pass at inference time—a design that balances learning depth with practical efficiency.

Performance Gains#

Testing on JsDeObsBench, the Coconut-based model substantially outperforms existing methods:

• 50% syntactic validity (code that is grammatically correct), compared to 15% for direct fine-tuning and 25% for zero-shot baselines
• 80% semantic correctness among valid outputs, indicating superior faithfulness to the original program logic

These results suggest the approach preserves both structural integrity and functional accuracy more reliably than simpler alternatives.