Skip to content

Security: aria-ml/dataeval-flow

Security

SECURITY.md

Security policy

This document describes how to report vulnerabilities in dataeval-flow and how the maintainers triage findings from automated scanners. It is the project's response to JATIC SDP requirement DSOR-1-H-3 (false positives must be dismissed with a documented justification).

Reporting a vulnerability

Do not open a public GitLab or GitHub issue for a security vulnerability. Email dataeval-flow@ariacoustics.com with:

  • A description of the issue and its impact
  • Steps to reproduce (a minimal config / dataset / command line is ideal)
  • Affected versions / commits / container tags
  • Your name and (optionally) a credit preference for the eventual advisory

You can expect:

  • An acknowledgment within 3 business days
  • A triage decision (accepted / false positive / duplicate) within 10 business days
  • A fix or mitigation plan within 30 business days for accepted Hard findings, tracked against the next v0.x.y release

For non-security bugs, follow the regular process in CONTRIBUTING.md.

Supported versions

Only the latest minor release on the main line receives security backports today. Once dataeval-flow reaches JATIC Maturity I (target v1.0.0, see ROADMAP.md), supported-version windows will widen per JATIC SDP RS-5-S-1.

Version Status Security backports
0.3.x (current alpha) active yes
anything older unsupported no

Automated scanner coverage

Every merge request and every main pipeline runs four security scans (see .gitlab-ci.yml):

Scanner Scope SDP requirement Suppression file / mechanism
Semgrep (SAST) source code DSOR-1-H-1 / DSOR-1-H-2 sast_excluded_paths input + per-rule nosem: comment
Gemnasium Python dependency manifest DSOR-2-H-1 / DSOR-2-H-2 GitLab vulnerability-report dismissal (with comment)
GitLab Container Scanning (Trivy analyzer) published container images DSOR-3-H-1 / DSOR-3-H-2 / CS-2-H-2 GitLab Vulnerability Report dismissal (with comment)
GitLab Secret Detection working tree DSOR-4-H-1 / DSOR-4-H-2 SECRET_DETECTION_EXCLUDED_PATHS on the job; no secrets

Findings flow into the GitLab security dashboard. Cosign attestation (cosign attest --type cyclonedx) is also published per image for downstream verification (CS-2-H-4).

Base image trust (CS-1-S-2)

All container base images referenced by this project, both the runtime variants in docker/ and the helper images used by CI jobs in .gitlab-ci.yml (python:slim, mambaorg/micromamba, ghcr.io/astral-sh/uv, docker:25.0.5-git, alpine:latest, etc.), are pulled through the program-owned JATIC GitLab registry / pull-through proxy. The registry is the trust boundary: images that are not on the program's approved list cannot be fetched by the runners, regardless of what a Dockerfile or CI job declares. CS-1-S-2 is therefore satisfied at the infrastructure layer, not the repo layer. No per-image attestation is maintained in this repository.

Dismissing a false positive

The same workflow applies whether the finding came from SAST, dependency scanning, container scanning, or secret detection.

  1. Verify it is a false positive. Read the rule / CVE description, inspect the code or dependency, and confirm the finding does not apply to how dataeval-flow uses the affected component. If you are unsure, treat it as a true positive and fix.

  2. Get a second opinion. Security suppressions land via a normal MR reviewed by a maintainer. The MR description must link to the finding (Semgrep rule ID, CVE, secret-rule ID) and to any upstream advisory or discussion that supports the dismissal.

  3. Record the suppression in the right place. The justification text must be specific to this project. Generic statements like "not exploitable in our case" without context will be rejected in review.

    • For Semgrep / SAST:

      • Per line, add # nosem: <rule-id> # Justification: <reason> on the triggering line.
      • Per file or per directory, add the path to the sast_excluded_paths input of the python-security component in .gitlab-ci.yml with an inline # <reason> comment above the change. It lists config, docker, docs, output and tests today.
      • Per analyzer, extend the sast_excluded_analyzers input only when a whole analyzer is out of scope. It excludes the analyzers for JavaScript, Go, PHP, Java, .NET, Elixir, Apex and Kubernetes manifests today, because the project ships only Python.
    • For Gemnasium / dependency scan:

      • The preferred fix is to bump the offending package in [tool.uv].constraint-dependencies (a constraint, not a runtime dep). See the existing CVE pins in pyproject.toml.
      • When a fix is not yet available upstream, dismiss the finding in the GitLab vulnerability report with a comment containing:
        1. CVE ID
        2. Why the project is not exploitable (for example, the affected API is never called, or the vulnerable code path requires untrusted input that does not reach it)
        3. A re-evaluation date (typically 90 days out)
    • For container scanning (the GitLab container_scanning job, one for each image variant, Trivy analyzer):

      • Suppressions live in the GitLab Vulnerability Report, not in a file in the repo. Open the finding, select "Dismiss", choose a reason and leave a comment containing:
        1. CVE ID
        2. Why the project is not exploitable, specific to how the affected component is used in our images. Generic statements will be rejected at audit.
        3. A re-evaluation date (90d for HIGH/CRITICAL, 180d for MEDIUM)
      • The severity gate is set via CS_SEVERITY_THRESHOLD in .gitlab-ci.yml. The container-scanning job's allow_failure: false ensures un-dismissed findings at or above that threshold fail the pipeline.
    • For Secret Detection:

      • True positives must be rotated immediately (the secret is already in git history, so assume it is compromised) and the commit purged with git filter-repo if it was added in the same MR.
      • False positives (test fixtures, public keys, example tokens) belong in tests/fixtures/ and the path added to SECRET_DETECTION_EXCLUDED_PATHS on the secret_detection job in .gitlab-ci.yml, with an inline justification. The job sets no exclusions today.
  4. Set an expiry. All non-CVE-fixable dismissals carry a re-evaluation date (90 days for HIGH/CRITICAL, 180 days for MEDIUM). When the date passes, the suppression must be removed or re-justified. There are no indefinite dismissals.

Audit and review

The set of active suppressions is reviewed at every release cut:

  • The release-prep checklist (see BRANCHING.md) requires walking the GitLab Vulnerability Report dismissal history (for both the dependency-scanning and container-scanning analyzers) and any in-line nosem: comments in the source to confirm each justification is still valid.
  • Suppressions whose re-evaluation date has passed without renewal block the release.
  • The aggregate count of active suppressions is reported in the release notes so it is visible to downstream consumers.

Hardcoded-secret policy

The project ships no hardcoded secrets (DSOR-4-H-2). The Cosign private key used to sign container images is supplied via the COSIGN_PRIVATE_KEY_B64 CI variable. Only the corresponding public key is committed (see docker/cosign.pub). If you discover what looks like a credential, password, token, or private key in this repository, even in a test fixture, report it via the channel above. Do not file a public issue.

There aren't any published security advisories