ZDICT trainFromBuffer is never told the length of the samples buffer. Before 0.4.10, zstd::dict::from continuous (initially train dict) called it directly: up to 0.2.1 without checking the sample sizes against the buffer at all, and from 0.3.0 with a sum that wraps around when overflow checks are off, as in release builds. Sizes that do not describe the buffer made the trainer read past its end.

This requires the sample sizes to come from an untrusted source. From 0.4.10, training goes through zstd-safe, where the same flaw is covered by the zstd-safe advisory for it, fixed in zstd-safe 8.2.0.