Reboot translates C interpreters to safe Rust with minimal human intervention

Reboot, introduced by Bo Wang and colleagues, addresses the challenge of translating C programs—especially interpreters handling untrusted input—to Rust's safe subset. The technique translates six interpreters (6k–23k lines of C) to safe Rust with minimal human help: each required only 1–11 brief interventions. All translations passed 100% of original test suites and 62–92% of newly created validation tests. A security case study on the mujs interpreter showed that memory vulnerabilities common in C—like heap buffer overflows and use-after-free—were eliminated in the Rust output.

Reboot's effectiveness stems from two key innovations. First, feature reduction breaks the translation into feature-based milestones, starting from a minimal yet testable version and incrementally restoring features, with validation at each step. Second, a multi-agent architecture orchestrates coding agents through automated validation and feedback, keeping long workflows on track with minimal human oversight. An ablation study confirmed that feature reduction boosts pass rates by 6–20% over a pure multi-agent approach, demonstrating its critical role in translation correctness.