ROCm security: scanning workflows, Dependabot and SECURITY.md

[gulsumgudukbay]

Description

Adds the ROCm GitHub org security baseline to this fork, as requested by the ROCm Security Baseline rollout. The org-wide "chore: add security scanning workflows" PR was opened on ROCm/jax, rocm-jax, TransformerEngine and jaxpp but never on this repo, so this adds the same files directly.

All files are copied verbatim from ROCm/rocm-repo-template at 80b94559e:

• .github/workflows/pr-security-scan.yml: scans each PR (changed files only) and reports to the job summary.
• .github/workflows/weekly-security-scan.yml: full scan every Saturday at 10:00 UTC, uploads SARIF to the Security tab.
• Both call the pinned ROCm/rocm-security-gh security-baseline.yml (v1.0.6): CodeQL, Bandit, Gitleaks, Trivy and zizmor, configured under .github/scan_tools_configs/.
• .github/dependabot.yml: daily, grouped version updates for GitHub Actions.
• SECURITY.md: the only foundational file this repo was missing. README.md, LICENSE, CONTRIBUTING.md, .gitignore and .github/CODEOWNERS already exist and are left identical to upstream, so they don't conflict on sync.

This is downstream-only and never goes upstream. Upstream has none of these paths, so it won't conflict when rocm-main is synced. It becomes a fourth ROCm commit replayed on each sync.

Please merge with "Rebase and merge" (or "Squash and merge"), not a merge commit, so rocm-main stays linear.

Notes for reviewers:

• Dependabot will propose action bumps across all workflows, including upstream's. Only take the ones that touch ROCm workflows; upstream-owned workflows should stay identical to upstream.
• Existing findings are almost all in upstream code. All 70 open Dependabot alerts are in upstream-owned requirements files (src/maxtext/inference/mlperf/requirements.txt and the TPU post-train requirements), which no ROCm job installs. The 6 secret-scanning alerts are AMD codename keyword matches inside upstream tokenizer vocab files and the C4 test dataset, so they are false positives.

Second commit (a6a2d47da) fixes every finding in ROCm-owned code. zizmor 1.30.0 and Trivy 0.70.0 (the versions the baseline uses) went from 24 findings to 0 and from 2 to 0:

• Workflow inputs, step outputs and github.token reach run: scripts through env: instead of ${{ }}, in the ROCm branch of run_tests_against_package.yml and in build_rocm_image.yml.
• actions/checkout is pinned to the v5.1.0 SHA upstream already uses. The TE wheel build no longer persists checkout credentials.
• The ROCm Dockerfile gets HEALTHCHECK NONE and an explicit USER root, with an inline trivy:ignore:DS-0002 stating why it stays root. Upstream Dockerfiles still report the finding.

The remaining Bandit, Trivy and Gitleaks failures are all in upstream-owned code and are left for upstream. Bandit and Trivy scan the whole tree on every PR. Gitleaks attributes three upstream files to this PR's commit only because of the shallow checkout.

Tests

• zizmor and Trivy run locally, before and after the fix commit, with the baseline's scanner versions.

• pre-commit run --files <the 8 files>: codespell, yamllint and actionlint pass.

• All workflow YAML and the gitleaks TOML parse.

• The PR scan itself runs on this PR.

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[github-actions]

Caution

3 potential secret finding(s) were committed in this pull request. Treat each credential as compromised and revoke or rotate it now, even if a later commit removed it. Deleting the line or force-pushing does not undo the leak: the commits stay reachable from this pull request and from every clone or fork, and automated scanners harvest secrets pushed to public repositories within minutes.

If a finding is a test fixture or a false positive, allowlist it in the configuration of the scanner that reported it.

Security scan found 3 critical or high finding(s) on lines changed by this pull request (coverage incomplete)

Scanner	Severity	Rule	Location	Finding
Gitleaks	HIGH	generic-api-key	.gemini/commands/maxtext-review.toml:139	Detected a Generic API Key, potentially exposing access to various services and sensitive operations.
Gitleaks	HIGH	generic-api-key	codecov.yml:28	Detected a Generic API Key, potentially exposing access to various services and sensitive operations.
Gitleaks	HIGH	generic-api-key	tests/assets/logits_generation/generate_dpo_golden_data_and_compare_pytorch_logits.py:49	Detected a Generic API Key, potentially exposing access to various services and sensitive operations.

Coverage was incomplete:

• Security scan / Security scanners concluded with failure.

View the security scan run for complete reports.

“New” means the finding is located in a changed file on an added or modified line; it is not a base-versus-head findings delta.