CVE-2026-18287

The 'load_' prefix in this library is a semantic trap. The load_time_series_segmentization_benchmark function does not do what its name implies — it executes arbitrary code from the loaded data via an eval() call, violating the foundational expectation that 'load_*' functions retrieve and deserialize data without side effects. This is not a missing input validation issue; it is a broken API contract that lures developers into treating untrusted benchmark files as safe data when they are, in fact, executable payloads. This matters critically because loading benchmark data is the library's primary workflow, not an edge case. Time-series researchers routinely exchange segmentation results through GitHub repositories, papers, and collaborative environments — each of these channels is now a potential injection vector. The CVSS 'user interaction required' metric obscures this: the interaction is opening a file, which is exactly what every benchmark-sharing workflow in this ecosystem demands. A researcher pulling a colleague's benchmark file to compare segmentation results has no reason to expect they are executing code. You must assume this pattern may not be isolated. Audit every function in the library prefixed with 'load_' for eval(), exec(), or similar code execution primitives — the absence of any security review history in the commit log suggests this was exploratory code that survived through institutional forgetting rather than active negligence. The fix is not just removing this eval() call; it is proving through the commit history that the entire serialization surface has been audited. Until the library provides explicit documentation that these functions execute code and publishes a security audit of its load_* family, treat every benchmark file from untrusted sources as potentially malicious. If you have already loaded benchmark data from this library in production or collaborative environments, treat those systems as potentially compromised — the attack surface is anyone who has ever shared a pickle or JSON benchmark file with a colleague, which is the entire user base.

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