Jailbreak uses LLMs to regenerate high-performance storage readers

Researchers present Jailbreak, an approach that bypasses database engines by using Large Language Models to regenerate operator-specific table reading components directly from storage files.

• The system leverages LLM-assisted code synthesis to decode complex database file formats specified by source code and documentation.
• It materializes data as in-memory Apache Arrow buffers consumable by query engines like DuckDB, Apache Spark, cuDF, and Spark RAPIDS.
• Validated against JDBC/ODBC baselines using the TPC-H benchmark, Jailbreak achieves up to 27x speedups in end-to-end analytical throughput.

This methodology breaks data lock-in across database systems by turning opaque formats into directly queryable artifacts without human-engineered parsing logic.