GitHub Security Lab

GitHub Security Lab offers an LLM‑driven autonomous fuzzing pipeline for C/C++ projects called the Fuzzing Taskflow. The taskflow identifies entry points, analyzes build systems, writes fuzz harnesses, runs AFL++, gathers coverage, and iteratively improves harnesses based on coverage feedback. It triages crashes and writes vulnerability reports. Built on the GitHub Security Lab Taskflow Agent, the pipeline uses a SQLite database (fuzz_context.db) for state, executes each harness twice (AFL‑instrumented and coverage‑instrumented binaries), and employs a time‑budget doubling loop with plateau detection to cease when further coverage gain falls below 1 %. Structure‑aware fuzzing is enabled through per‑format dictionaries, autogenerated source‑level mutators, coverage‑driven dictionary enrichment, and a corpus‑splice operator. The default model is Claude Sonnet 5; the framework can be configured to use alternative models.· Graphiti 生成的实体摘要