Four zero bytes at byte 14 walk FFmpeg's VPK demuxer into a division by zero

According to the FFmpeg issue, `vpk_read_header` does validate that the channel count is positive, and `vpk_read_packet` divides by that count as zero anyway ^[5][6]. The report explains the gap by pointing at the harness: in the fuzzer's custom-I/O setup, the bytes seen during probing and the bytes seen during packet reading can diverge ^[5]. That clause is load bearing, and it sits awkwardly beside the crash metadata, which records the input as fully deterministic with no preconditions and no network ^[10], and beside the claim that any FFmpeg-based application that opens the file and reads a packet dies ^[1].

For the 21 bytes to transfer to a stock build reading a file off disk, the short read has to land the same way. The demuxer parses a 24-byte header out of a 21-byte file, three bytes it never had ^[3]. The four zeros sit at bytes 14 through 17, inside the file ^[5]. The remainder of that header comes from whatever the I/O layer returns past the end, and the I/O layer is precisely the piece the fuzzer replaced. The article's author says he cloned the repo, attempted the crash on his own Ubuntu box, and flagged the one place his results diverged from the report ^[20]. That flag is where I would start.

Then the discovery cost. Ten hours forty-three minutes is 38,580 seconds, so 495,211 executions is roughly 12.8 per second ^[1]. That is slow, and the fuzzer's README says so, conceding that AFL remains the better choice for production fuzzing at scale ^[18]. A 13,188-entry corpus against 495,211 executions is a 2.7 percent retention rate ^[2]. The scheduling earned this finding rather than the execution rate: 147 mutation operators across nine categories, with 14 scheduler modules under Elo arbitration ^[17], aimed at a demuxer for Sony PS2 audio files ^[2]. VPK support is presumably load-bearing for someone.

The craft is worth naming. The repo carries an `AGENTS.md` that the human maintains for the agents working on the code, with rules such as always find the closest existing example and match its convention ^[19]. A README that publishes its own throughput disadvantage ^[18] tells you more than most benchmark tables do.

Adoption cost on the defender side is small: a guard on `nb_channels <= 0` at the top of `vpk_read_packet`, an `AVERROR_INVALIDDATA` return, and a regression test ^[12]. Exposure cost depends on your process model. A Medium-rated denial of service ^[11] is a restart for a transcode worker that runs out of process, and an outage for anything that links FFmpeg in-process and parses uploads it did not produce. Since the bug class was already characterised on ffmpeg-devel ^[13], the remaining work is not research. It is inventory: every call site that divides by a count a header handed it.