Only the latest 0.x release on PyPI receives security updates. Patches land in a new 0.x.y release; older versions are not backported.

Do not open a public GitHub issue for security reports.

Email the maintainer at david@villacis.net with:

• A description of the vulnerability and its impact.
• Reproduction steps (a minimal failing case is ideal — the simpler, the faster we can confirm and patch).
• The affected version (pip show sparho or sparho.__version__).
• Your preferred attribution in the resulting advisory, or anonymous.

Expected response: acknowledgement within 5 business days; an initial triage assessment within 14 days. We follow a 90-day coordinated disclosure window — if the fix lands sooner the advisory is published on the release date, otherwise on day 90 regardless.

sparho runs locally inside a user's Python process and consumes in-memory numpy / scipy data structures. The realistic attack surface is:

• FFI boundary. Rust kernels validate every caller-supplied i32 index against the matching slice length before any as usize cast (v0.3.1 hardening). Malformed scipy.sparse.csc_matrix input produces a ValueError, not a Rust panic — the release profile sets panic = "abort" as the safety net for any remaining bug.
• Dependency advisories. A cargo-audit job runs on every PR; the one current ignore (PyO3 0.22 PyString::from_object — RUSTSEC-2025-0020) is tracked on the ROADMAP behind a pyo3 → 0.24 bump. Python deps are checked transitively via the floor-version ci-min-deps matrix job.

Out of scope (these are the user's responsibility):

• Loading untrusted pickle artifacts. SearchResult and the wrapper estimators are picklable, but pickle is unsafe with attacker- controlled input by construction.
• Adversarial training data. The library makes no robustness guarantees against data poisoning or membership inference.
• Side-channel attacks on the inner solver (timing, cache).

None yet. Be the first.