The work explores why ordinary coverage-guided fuzzers struggle to reach deep code in GStreamer’s MP4 demuxer. MP4 files are nested trees of atoms whose size fields must remain consistent with their contents and ancestors; random insertions or deletions often make a file invalid, so parsing stops before reaching complex routines. A fully structure-aware fuzzer would be costly to build and could also miss bugs that depend on malformed inputs. Before designing a targeted approach, the researcher established that OSS-Fuzz had no MP4 coverage because its GStreamer targets did not include the Good plugin set, which contains the MP4 parser. A locally enabled, general-purpose discovery fuzzer ran for 16-24 CPU-weeks.
That baseline uncovered five memory-safety bugs: two heap out-of-bounds reads in MP4 parsing, plus denial-of-service flaws in TMPlayer and SubRip subtitle handling. The subtitle findings arose because format-agnostic mutations can drift into other supported file types, an unwanted result when MP4 depth is the goal. To focus mutations, the researcher built a GStreamer pipeline that routes inputs specifically through qtdemux, causing non-MP4 files to fail early and receive little further attention from the fuzzer. This pipeline was faster than the general-purpose target and found an additional heap out-of-bounds read in PlayReady DRM metadata parsing (CVE-2026-46471), which could expose memory contents or help bypass ASLR.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.