All notable changes to the workflows in this repository will be documented in this file.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning. See README.md for the SHA-pinning contract that consumers are expected to follow.
0.6.0 - 2026-11-13
Secret-scan maintenance release — no new scans, inputs, or hooks. One consumer-visible change: the Pulse secret-scan job is renamed, so its status check becomes … / pulse-secret. A repository whose branch protection requires the old … / pulse check must update that rule in the same change that repins, or the rule waits on a check that never reports. Inputs, permissions, failure_policy behaviour, and every pre-commit hook are unchanged.
- Secret scan — the job is renamed from
pulsetopulse-secret, matchingpulse-ossandpulse-malware. Its status check becomes… / pulse-secret, so repositories that require the old… / pulsecheck must update branch protection when they repin.
- Secret scan — the filesystem scan checks out at depth 1 instead of fetching every branch and tag.
filesystemmode reads the checked-out tree, not packed Git history, so the full fetch bought no scan coverage. InNVIDIA/cuda-pythonit spent ~317s of a ~404s job downloading 3.54 GB — most of it the unrelatedgh-pagesbranch — to scan 33 MB of source.
0.5.0 - 2026-11-03
Additive release: two reusable workflows (vulnerability, malware) and one pre-commit hook (GuardWords). Existing suite callers need no change — enable-vuln-scan and enable-malware-scan default to false, and the scans' permissions (contents: read + id-token: write for the vulnerability scan, id-token: write alone for the malware scan) are already inside the suite's permission union.
- Vulnerability scan (SCA) reusable workflow —
.github/workflows/vuln-scan-pulse-oss.yml. Runs NVIDIA's Pulse OSS scanner against a disposable copy of the checkout and reports to the nSpect backend over a Charon loopback tunnel. Consumers supply no credentials: GitHub OIDC is the only identity the scan presents, and thenvcr.iopull login comes from Vault. The scanner image and the tunnel settings are owned by the workflow, so a caller can substitute neither.pull_request/pull_request_targetare rejected. Inputs and prerequisites are in the workflow catalogue; build-environment setup is in the scan guide. - Vulnerability scan — scan targeting.
nspect_idis optional: left empty the scan runs unregistered, which is the only lane open to a repository with no nSpect registration and is rapid-only.sbom_nameselects the registration when the registered branch carries more than one SBOM entry, andinventory_writebackrequires it.copy-pr-botmirrors (refs/heads/pull-request/*) are rapid-only. - Vulnerability scan — findings never decide the scanner's exit code. The CLI always runs with
--no-failandfailure_policyis applied to the validated report, so a non-zero exit is always a failed scan rather than a finding. - Vulnerability scan — optional build controls, off by default and forwarded by the suite: prepared Python wheels, typed Detect properties, and redacted console/debug logging. Snippet scanning and snippet-match export are deferred to a follow-up release.
- Malware scan reusable workflow —
.github/workflows/malware-scan-pulse.yml. Asks the Pulse Malware Service to scan the calling run's own commit with ReversingLabs over the production Charon loopback tunnel. The service fetches the commit itself, so nothing is checked out and the scan cannot be pointed at another repository or commit.failure_policyisstrict(default — SUSPICIOUS blocks too) ormalicious-only; a MALICIOUS verdict, and any scan that ends without one, block under both.pull_request/pull_request_targetare rejected. Prerequisites are in the workflow catalogue. - Security suite —
enable-vuln-scanandenable-malware-scan, each with its ownvuln-*/malware-*inputs, plus an unprefixednspect-id: an nSpect ID identifies the repository rather than one scan, so every scan reporting to nSpect reads the same one. The malware scan requires it — enabling that scan without it fails the preflight. - GuardWords (pre-commit) —
guardwords-scan-code-scanningsends the staged diff to NVIDIA's internal code-scanning service and blocks the commit on blocking findings. It is for NVIDIA employees only: where the client or service is unavailable it skips with a notice, so external contributors are never blocked.hooks/guardwords/install_client.pysets up the client in one command. See the GuardWords hook guide. - Pre-commit hook catalogue —
hooks/README.md, the hook-side counterpart to the workflow catalogue. - CI — contract tests covering every gate in both new workflows, plus live integration runs. The vulnerability scan's integration is ungated, so its Vault and Charon prerequisites stay hard requirements.
- Third-party notices —
teleport-actions/setupandactions/upload-artifact, both pinned by SHA.
- Security suite — the preflight reports every missing companion input in one run rather than exiting on the first.
- Pin policy — the recommended pre-commit
rev:pin is now the 40-character commit SHA with the release tag as a# frozen:comment, the formpre-commit autoupdate --freezewrites, matching the SHA pin already required for reusable workflows. Existing release-tag pins keep working; switch withpre-commit autoupdate --freezeat the next refresh. See Pin policy per surface.
- Charon is a prerequisite for both scans. The tunnel authorizes per repository and per ref, so a consuming tenant must grant the Pulse backends and list every ref it scans from. Gating pull requests needs the
pull-request/*mirrors listed; without them both scans are post-merge and scheduled gates. The malware scan-submission route is not granted by default and needs platform approval. - Scanner upgrades ship as workflow releases. Both scanners are pinned by digest — Pulse OSS
0.5.3-charon.11, Pulse malware1.3.0— so a new scanner reaches callers only through a new release of this repository. - Vulnerability scan — SARIF publication is a follow-up. Findings land in the job summary and an optional artifact, not the Security tab.
- Vulnerability scan —
build: trueruns the repository's build inside the scanner. Maven and Go graphs resolve only through a build. Prefer it on post-merge and scheduled runs; see the workflow catalogue before enabling it on pull-request mirrors. - Vulnerability scan — Persistent and write-back scans are pilot-only. Inventory write-back with
sbom_nameagainst a branch with several SBOM entries has not been validated end-to-end. - Malware scan — Detection is not yet proven end-to-end. CI asserts transport and a CLEAN verdict; proving a blocking verdict needs a ProdSec-approved detection sample, and the contract tests cover how one is handled.
- GuardWords — No CI counterpart. The hook is the only surface and it skips wherever the client or service is unavailable, so a commit that skipped is not re-scanned before merge.
0.4.0 - 2026-08-19
Additive release. Callers passing languages: '["actions","python"]' need no change; the declared permissions: are unchanged, and the analyze job keeps the name CodeQL (<language>) so required status checks continue to match.
- SAST (CodeQL) — build mode is resolved per language instead of one value applied to every matrix leg, so sets with conflicting requirements work in a single call.
'["go","rust"]'could not previously be expressed (goneedsautobuild,rustsupports onlynone). Substitutions are reported as job-log notices and the resolved matrix is written to the job summary. - SAST (CodeQL) —
languagesentries may be objects carrying a per-languagebuild-modeorruns-on:'[{"language":"swift","runs-on":"macos-14"},"rust"]'. Per-entryruns-onmakesswiftusable, as it requires a macOS runner. - SAST (CodeQL) — a
resolvejob validates the language set before any analysis starts, rejecting unknown languages,manualmode, unsupported explicit modes, duplicate languages, malformed JSON, and empty lists. Newresolve-runs-oninput (defaultubuntu-latest); the suite passessuite-runs-onthrough to it. - Documentation — the workflow catalogue now lists the accepted
languagesvalues with per-language buildless support, plus the two coverage gaps: no shell analyzer (useshellcheck) and no CUDA analyzer (c-cppskips.cu/.cuh).
- SAST (CodeQL) —
build-modeis now a preference applied where the language supports it; an explicit per-language mode is exact and errors if CodeQL rejects it. No existing caller changes behaviour. - SAST (CodeQL) — the
analyzejob keeps the nameCodeQL (<language>), so required status checks still match. A new non-required check,Resolve language matrix, appears alongside it.
- Documentation —
build-modewas described as working "for every language":goandswiftrejectnone, Kotlin requiresautobuild, andrustsupports onlynone. - Documentation — the CodeQL prerequisite said Default setup must be off entirely. The conflict is per language: both publish
/language:<lang>, so only an overlap collides. Default setup on other languages can coexist, as it already does on this repository. - Documentation — the catalogue gave CodeQL's
runs-ondefault aslinux-amd64-cpu4(actual:ubuntu-latest) and required annv-gha-runnerslabel for every workflow, which does not apply to CodeQL.
0.3.0 - 2026-08-11
- Security suite reusable workflow —
.github/workflows/security-suite.ymlruns the scans a caller enables (enable-secret-scan,enable-sast-scan) in parallel behind one pinned reference. Every scan is opt-in, including scans added in future releases, so onboarding a repository to the suite never turns on a scan it did not ask for; category-prefixed inputs preserve the child workflows' runner and policy configuration. A suite call that enables nothing warns instead of failing. - SAST (CodeQL) reusable workflow —
.github/workflows/sast-scan-codeql.yml. Generic, matrix-driven CodeQL analysis (languages,build-mode,queries,packs,config-file,runs-oninputs) with thegithub/codeql-actionsteps pinned by SHA. Positioned as the customization-tier lever — the fleet baseline for CodeQL is GitHub Default setup via org/enterprise Security Configurations. See.github/workflows/README.mdfor prerequisites (GHAS/code scanning, Default-setup conflict) and usage.
- Secret-scan (Pulse) — the reusable workflow no longer uploads SARIF on CI self-test runs (
ci_test_setup: true), so the disposable RSA fixture no longer publishes a code-scanning alert to the default-branch Security tab. Real consumer scans are unaffected. - Secret-scan (pre-commit) —
secret-scan-trufflehognow installs the SHA-256-pinned TruffleHog release in pre-commit's isolated Python environment rather than the per-user cache.
- Secret-scan (pre-commit) —
secret-scan-trufflehogno longer fails on pytest function names. A Lob API key istest_followed by 35 characters, so TruffleHog's Lob detector matched names such astest_gpu_conf_compute_attestation_reportand its verifier reported them as verified, the one result class the hook blocks on. Lob is now excluded from the hook's detector set.
0.2.0 - 2026-07-24
- Secret-scan (pre-commit) — the
secret-scan-trufflehoghook is now self-installing: on first use it downloads a pinned, checksum-verified TruffleHog release (3.95.9) into a per-user cache and reuses it thereafter, so contributors no longer installtrufflehogmanually. The archive is fetched over HTTPS and verified against a per-platform SHA-256 pinned from the release's cosign-signedchecksums.txt; the hook fails closed on any mismatch. Supports Linux/macOS/Windows (amd64/arm64).
0.1.0 - 2026-07-22
First pilot release. Ships the secret-scan surface (Pulse reusable workflow + TruffleHog pre-commit hook). Pre-1.0: interfaces may change in a minor release while surfaces stabilize with ProdSec.
-
Repository scaffolding —
README,ROADMAP,GOVERNANCE,CONTRIBUTING,SUPPORT,MAINTAINERS,SECURITY,CODE_OF_CONDUCT,LICENSE. -
Documentation — top-level
README.mdcovers the two-surface model (reusable workflow + pre-commit hook) with consumption examples and the pin-policy table;.github/workflows/README.mdis the workflow catalogue;ROADMAP.mdtracks per-scan status. -
OSRB compliance artifacts —
LICENSE,NOTICE,THIRD_PARTY_NOTICES.md, no-contributionsCONTRIBUTING.mdper OSRB template (maintainer PR workflow lives inGOVERNANCE.md), and source/config SPDX copyright headers. -
Contribution posture — clarified the OSS Type 2 model: public source code with NVIDIA-internal development and no external contribution intake.
-
Pin policy (
README.md→ Pin policy per surface) — 40-character commit SHA for GHAuses:; release tag for pre-commitrev:; branch references not acceptable. Release tags cut on consumer-visible behaviour or contract changes. -
Repository tooling — pre-commit hygiene configuration (
.pre-commit-config.yaml) runningpre-commit-hooks,yamllint,shellcheck,actionlint, andcodespell; agent guidance (AGENTS.md,CLAUDE.md). -
copy-pr-bot —
.github/copy-pr-bot.yamlto gate external-contributor pull-request CI on self-hosted runners. -
Secret-scan (OSS) pre-commit hook —
.pre-commit-hooks.yaml+hooks/trufflehog.sh, composing upstreamtrufflesecurity/trufflehog. Thesecret-scan-trufflehoghook scans changed files (pass_filenames: true) at thepre-commitandpre-pushstages, fail-closed with NVIDIA default--results=verifiedand--no-updateto keep the local scan fast. The wrapper resolves thetrufflehogbinary from$PATHand prints install guidance if it is missing. This is the local-advisory lane. -
Dogfooding —
.pre-commit-config.yamlruns thesecret-scan-trufflehoghook fromrepo: localso edits to the wrapper are self-tested; skipped onpre-commit.ci(notrufflehogbinary in that sandbox). -
Secret-scan (Pulse) reusable workflow —
.github/workflows/secret-scan-pulse.yml. Runs the Pulse Secret Scanner as the CI enforcement lane and publishes redacted SARIF (verified/unknown/unverified) to the repository Security tab. Enforcement via thefailure_policyinput (unverified|strict|all, defaultunverified; names match the GitLab secret-scan component). Fails closed on scanner/infra errors. -
Continuous integration —
.github/workflows/ci.yml. Lints all workflow/hook files and runs positive/negative Pulse integration tests across all threefailure_policyvalues. -
Third-party notices — TruffleHog (AGPL-3.0, developer-installed CLI) and the direct workflow actions (
actions/checkout,hashicorp/vault-action,docker/login-action,actions/setup-python,github/codeql-action/upload-sarif) recorded inTHIRD_PARTY_NOTICES.mdwith SHA pins where applicable.