Committed dependency audits

Dependency Lock Audit runs on every push and on every pull request to main. Its dependency-lock-audit aggregate succeeds only when all six ecosystem jobs succeed. Branch protection must require this aggregate after its first hosted run has reported; a workflow file alone does not establish that setting.

The inventory comes from checked git ls-files output. Each job records every lock digest before and after execution. New locks in a supported format enter the audit automatically; an unsupported additional .lock format fails the inventory. Project dependencies are neither installed nor rewritten by audits.

Ecosystem	Committed input	Advisory path
Python	Every hashlocked requirements/*.txt profile	Strict pip-audit queries for every pinned version and marker branch
Rust	Every Cargo.lock	cargo audit --file … --format json --deny warnings --url https://github.com/RustSec/advisory-db
npm	Every package-lock.json	Exact registry identity validation, then npm audit --package-lock-only --include=dev --include=optional --include=peer --registry=https://registry.npmjs.org --json
Go	Every go.mod and associated go.sum	OSV queries for standard library, module pins and checksum-only versions
Julia	Every format-2.0 or format-2.1 Manifest.toml	OSV queries for every exact package version, including build suffixes
Pixi	Every pixi.lock	Pinned pixi-audit, with complete JSON coverage validation

Run an ecosystem without running a test suite:

Bash

python -m tools.security_scan.locked_dependency_audit \
  --ecosystem julia --output-dir /path/to/new-audit-packet

Findings of every severity, ignored or unchecked packages, malformed responses, database errors, tool failures, timeouts and changed input bytes block the job. The direct OSV adapter retains each request and response page, including HTTP error bodies. Per-page receipts bind the request and response digests and HTTP status. Connection failures retain their exception types, without exception text or headers. Responses remain bounded to 16 MiB plus one detection byte; each request has a 30-second timeout. Go call analysis does not filter advisory results. Julia package versions are not converted to Python or generic upstream versions.

Julia 1.13 writes format 2.1 manifests with registry metadata. Both supported formats retain every dependency entry; path, Git and unversioned dependencies remain refused. The native manifest regression resolves the maintained FFTW project with the installed Julia 1.13 runtime and compares every generated package/version identity with the audit queries.

npm silently omits unversioned packages from advisory submission. Before running it, the adapter validates every installation entry in a version-2 or version-3 lock, retaining its path, registry name and exact semantic version. Linked, unversioned, nonregistry and empty dependency inventories fail. These require a reviewed advisory adapter before they can enter a passing audit. The response must cover the complete validated package count, with integer counters for all severities. A zero exit status or a clean-looking total alone is insufficient.

The Studio lock selects source-map-js 1.2.2 for GHSA-68fv-2mgg-jv7q. The update retains the remaining dependency graph. Its npm audit still includes development, optional and peer dependencies; a passing npm result does not qualify the other ecosystem jobs or the whole security packet.

The installed native-wheel profile is generated from requirements/wheels-test.in with the development lock as its constraint. Its universal resolution retains the Python 3.10 and 3.11 NumPy branches and the platform-specific pytest dependencies. Regenerate it with:

Bash

uv pip compile --universal --python-version 3.10 --generate-hashes \
  --constraint requirements/ci-dev.txt \
  --output-file requirements/wheels-test.txt requirements/wheels-test.in

This profile includes the framework and shared-fixture dependency used by the installed batch-encoding contracts. It remains part of the blocking Python lock audit, including every marker branch.

Rust report validation requires explicit unfiltered settings, all supported informational warning categories, consistent typed finding counters and zero warnings. Missing settings cannot stand in for an unsuppressed audit. Effective project or Cargo-home audit configuration is digest-bound before and after scanning. Disabled database fetch, stale acceptance and database origin or cache overrides fail before the scanner runs. A custom advisory database requires a separately reviewed custody adapter.

The reviewed Python CPU profile preserves torch's +cpu lock identity, download origin and hashes while querying the upstream public version. Its report explicitly records both identities. Other local version suffixes are refused. This checks upstream package advisories; it does not attest that a particular CPU wheel has no build-specific vulnerability.

The current Pixi beta covers conda-forge and PyPI. Modular packages reported as unchecked cannot establish a passing complete-lock audit. A Pixi lock without a readable package record list is refused before the scanner starts. No blanket waiver or automatic --fix is applied. Advisory remediation needs reviewed fixed versions and ordinary integration validation before changing the lock.

The additional security-packet OSV runner also treats every nonzero scanner exit as failure. A partially written report cannot turn a service error green. Audit evidence does not itself establish a release, hosted CI success or eligibility to delete failed Actions history.