Skip to content

Security: xaedro-ai/hakuli

SECURITY.md

Security Policy

Supported versions

The latest released minor version receives security fixes.

Version Supported
2.x Yes
1.x No

PyTorch dependency advisories

Hakuli 2.x requires a patched PyTorch installation. New installations must use torch>=2.15.0 with Python 3.11+.

Advisory Impact Fixed in PyTorch
CVE-2025-32434 Code execution via malicious checkpoint 2.6.0
CVE-2026-24747 Code execution via malicious checkpoint 2.10.0
GHSA-v32g-mr63-3j83 File disclosure via malicious checkpoint 2.15.0

These advisories affect upstream PyTorch checkpoint loading, not Hakuli's optimizer implementation. Do not load checkpoints from untrusted sources.

All benchmark installation entry points, including the Dockerfile and the legacy requirements-repro.txt command, use maintained dependencies. The vulnerable versions used for the October 2026 measurements are recorded only as provenance in benchmarks/historical-environment.md; new runs do not recreate that historical environment.

Reporting a vulnerability

Please do not report security vulnerabilities through public issues.

Use GitHub's private vulnerability reporting on this repository ("Security" tab → "Report a vulnerability"). Include:

  • a description of the issue and its impact,
  • a minimal reproduction (code snippet, configuration),
  • the affected version(s) and environment.

You can expect an acknowledgement within a few business days. We will keep you informed of the remediation progress and credit you in the release notes unless you prefer otherwise.

Scope

Hakuli is a training-time optimizer library. Relevant security concerns include, for example, unsafe deserialization paths, code execution during checkpoint loading, or dependency vulnerabilities. Model quality and numerical behaviour are tracked as regular bugs, not security issues.

There aren't any published security advisories