GHSA-w5x2-6474-pqp4

Affected versions passed caller-supplied bytes to calamine's Xlsx::new without first checking that they were a ZIP archive. Xlsx::new tests for password protection on its first line, which parses the input as an OLE/CFB container, and a sector-count field read from the file's own header reaches Vec::with_capacity without being checked against the file's actual length. A 512-byte input can therefore request several gigabytes; 9,261,285,372 bytes was measured.

The oversized allocation is always attempted. Whether it aborts depends on what the allocator can satisfy. On a large host with overcommit the reservation is granted untouched and the call returns an ordinary "not an xlsx file" error, which looks like a malformed file being correctly rejected. Under a memory limit the same bytes give memory allocation of N bytes failed and the process aborts, which is not a Result a caller can handle. Containers with a memory limit, small hosts and CI runners are where this lands, so testing on a development machine can wrongly suggest the crate is unaffected.

Every entry point reaches it, not only compare_bytes: the path- and reader-based APIs funnel through the same internal open. Neither Limits::default() nor Limits::hardened() prevents it, because max_input_bytes bounds the length of the input while the allocation's size comes from a field inside it.

Fixed in 3.2.0, which declines input that does not begin with the ZIP magic before the parser sees it, removing the path rather than bounding it. Callers who cannot upgrade can apply the same check before calling this crate.

The underlying defect is in calamine, reported there independently as tafia/calamine#714 and unfixed at the time of writing. It is reachable from any crate that opens untrusted bytes with Xlsx::new.