Skip to content

SECURITY.md

Latest commit

 

History

History
67 lines (51 loc) · 3.05 KB

SECURITY.md

File metadata and controls

67 lines (51 loc) · 3.05 KB
Raw

Security policy

Last reviewed: 2026-07-21.

Supported versions

Security fixes are applied to the current main branch and included in the next release. Historical research releases remain immutable so their published artifacts and environments can still be reproduced; they do not receive backports.

Reporting a vulnerability

Please report suspected vulnerabilities privately through GitHub Security Advisories. Include the affected entry point, reproduction steps, expected impact, and any suggested mitigation. Do not open a public issue for an unpatched vulnerability or include live credentials, private data, or exploit payloads in a report.

General hardening suggestions that do not disclose a vulnerability may use the normal issue tracker.

Security boundary

This repository is a research pipeline operated by a trusted user. Its primary entry points are local CLI commands that process operator-selected datasets, models, configurations, and filesystem paths. The project does not provide multi-user isolation or sandbox arbitrary model and dataset artifacts.

The optional mgq-serve FastAPI wrapper is intended for local demos and integration tests. It binds to 127.0.0.1 by default and rejects non-loopback hosts unless the operator explicitly passes --allow-remote. That override does not add authentication, TLS, authorization, request quotas, or rate limiting; deployments outside a trusted host must supply those controls separately.

External artifacts and network access

  • Hugging Face libraries download model and metric artifacts selected by the configuration or CLI. Use pinned model revisions for controlled experiments.
  • ir_datasets and dataset loaders retrieve benchmark corpora and judgments from their configured upstream sources.
  • Release-reproduction commands download repository-owned result bundles and verify the archive and member hashes before consuming them.

Treat model weights, datasets, run files, JSONL inputs, and archives from untrusted sources as untrusted data. Run third-party artifacts in an isolated environment when their provenance is uncertain.

Repository controls

  • GitHub secret scanning and push protection cover repository history and incoming pushes.
  • A repository-local high-confidence scanner checks tracked text files in CI.
  • CodeQL analyzes Python changes and the default branch.
  • GitHub Actions are pinned to full commit SHAs and receive read-only repository contents by default.
  • Dependency updates are proposed by Dependabot. The direct-dependency snapshot is audited during security reviews and checked against the artifact registry.

Known limitations

  • requirements-lock.txt pins direct dependencies but is not a fully resolved, hash-locked transitive environment. Use isolated environments and review resolver output for sensitive deployments.
  • Downloaded models and datasets inherit the trust and availability properties of their upstream registries.
  • The local FastAPI wrapper is not a production-ready public API.