CVE-2026-18286

The CVSS 7.8 score assigned to CVE-2026-18286 overstates the practical risk by treating this as a direct attack vector when the exploitation path is fundamentally different. This vulnerability lives in `load_human_activity_segmentation_datasets`, a data loading function within the aeon library — an attacker cannot reach it through network adjacency or an exposed API. They must instead deliver a malicious dataset file to a victim and convince them to load it through this function. The user interaction requirement isn't a footnote; it's the entire precondition. This matters because the library occupies a trust boundary that it shouldn't. Data loading functions are where you process external, potentially untrusted data — if that pipeline contains dynamic code execution, the library has inverted its own security model. The EPSS score of 0.00342 reflects what practitioners observe: exploit probability stays low because the chain requires social engineering plus technical execution, not because the vulnerability is minor. For defenders, the priority actions are: audit your code for any calls to `load_human_activity_segmentation_datasets` or similar dataset loading functions and verify they aren't processing untrusted files; treat data loading libraries as untrusted boundaries regardless of their origin — don't assume a published library is safe simply because it's widely used; and scan your data pipelines for patterns like `exec()`, `eval()`, or dynamic string interpolation in code paths that handle external data. In research environments where datasets circulate through collaboration graphs, a single poisoned file can propagate through co-authors and papers, creating contamination that outlasts the technical vulnerability. The real exposure isn't the CVSS vector — it's the assumption that loading a dataset is a boring, safe operation.

Reviewed through automated stages and approved by a human before publication.