SandBench

Benchmark untrusted code in a throwaway sandbox.

Submit untrusted code. Get a verdict, a score, and an automatic repair — all in a hardened sandbox.

Core Capabilities

One secure pipeline, from evaluation to repair

Hardened sandbox verdict

Every run gets executed inside an ephemeral, hardened container. We drop capabilities, enforce read-only roots, set CPU/Memory limits, and monitor for egress or escape attempts.

Real hidden-test scoring

Submissions that complete successfully are scored against a private, hidden test suite. Get an accurate score represented as a fraction of tests passed.

LLM review & patch suggestions

Clean but failing submissions are automatically reviewed by an LLM model, producing suggestions and a concrete patch diff to fix the bug.

RL repair-and-rescore loop

Our repair pipeline automatically applies the suggested patches, re-runs the code in the sandbox, and computes the exact reward/improvement score.

Live leaderboard & playground

Deduplicated scores are posted to a live leaderboard. Visualize repair trajectories, step-by-step rollout animations, and cumulative rewards in the RL playground.

Docker + seccomp

Sandbox isolation

Hidden test suite

Scoring method

LLM + RL reward

Repair loop

Real-time

Leaderboard