Releases · alibaba/last-secure-code-benchmark

First public release of the Last Secure Code Benchmark (LSCBench).

tasks/: 450 Harbor tasks built from 150 vulnerability scenarios (115 from real CVEs, 35 expert-written Java applications), each at function, file, and repository granularity, covering 7 languages and 24 CWE classes. The same tasks are on Harbor Hub as last-secure-code-benchmark/last-secure-code-benchmark@v1.0.
traces/: the 2,700 agent runs (six models on all 450 tasks) behind the leaderboard.
Task images: lscbench/last-secure-code-bench on Docker Hub.

pip install harbor
harbor run -d last-secure-code-benchmark/last-secure-code-benchmark@v1.0 -a oracle

Navigation Menu