Skip to content

Supply evidence later slices: vulnerability/SBOM/posture/attestation adapters and selected rebuild-lab #45

Open

Description

@PhysShell

Classification

  • Change type: enhancement / deferred evidence adapters + research lab
  • Priority: P2
  • Expected value: medium-high after slice 1 proves the evidence/profile architecture
  • Implementation risk: high if started too early — can easily turn into a broad supply-chain scanner or unreliable rebuild farm
  • Repository owner: OwnAudit implementation/research layer
  • Depends on: OwnAudit Supply evidence slice 1: SourceLink/PDB and NuGet resolution-policy adapters #44 and PhysShell/Own.NET#271
  • Production status: deferred; do not start before the SourceLink/PDB + NuGet policy slice is stable

Problem

There are many useful supply-chain tools, but OwnAudit should not duplicate them:

NuGet Audit / OSV-Scanner / Grype     -> known vulnerability evidence
Syft                                  -> SBOM evidence
OpenSSF Scorecard                     -> repository posture evidence
SLSA / GitHub artifact attestations / Sigstore -> build/provenance evidence
rebuild experiments                   -> selected source-to-artifact correspondence evidence

The gap worth owning is not "another scanner". The gap is a .NET/NuGet/legacy-enterprise evidence aggregator that can combine those signals with SourceLink/PDB metadata, package policy and existing audit findings.

Goal

After the first supply evidence slice is stable, add optional adapters for mature tools and a separate selected rebuild lab.

The adapters must remain independent modules feeding the same evidence contract. They must not add a second report pipeline or replace existing tools.

Proposed later adapters

known-vuln

Inputs:

packages.lock.json
project.assets.json
NuGet package graph
OSV-Scanner output
NuGet Audit output
Grype output where SBOM exists

Rules:

SC007 VulnerablePackage

Non-goal: OwnAudit does not maintain a vulnerability database.

sbom

Use Syft or another external SBOM generator. OwnAudit consumes the result as evidence.

Questions to answer:

what packages are inside this artifact?
which transitive dependencies exist?
which binaries/native blobs are present?
which packages lack source/provenance metadata?

Non-goal: OwnAudit does not invent an SBOM format.

repo-posture

Use OpenSSF Scorecard or equivalent repo-health tooling for dependency source repositories.

Signals:

branch protection
pinned dependencies
binary artifacts
signed releases
security policy
maintained/stale status

This is repository posture evidence, not package proof. A good Scorecard result does not prove that a .nupkg matches a source commit.

attestations

Consume available build provenance evidence:

SLSA provenance
GitHub artifact attestations
Sigstore/cosign where applicable
in-toto/DSSE-style attestations

Non-goal: no local trust policy engine yet unless the evidence contract proves it is needed.

rebuild-lab

Selected source-to-artifact correspondence experiments only.

Candidate targets:

Internal.*
Sector.*
high-risk dependencies
packages with native binaries
packages touched by an incident
packages with missing/contradictory provenance evidence

Rules:

SC013 ArtifactNotReproducible
SC014 NupkgContainsUnexpectedBinary
SC015 PackageBuiltFromDirtyOrUnknownSource
SC020 RebuildAttempted
SC021 DeterministicRebuildMismatch
SC022 NormalizedDllMismatch
SC023 NupkgContentMismatch

This must not attempt to rebuild the NuGet ecosystem.

Profiles

Add only as opt-in profiles:

incident-response:
  enabled:
    - nuget-integrity
    - known-vuln
    - sbom
    - repo-posture
    - attestations
    - rebuild-lab

rebuild-lab:
  enabled:
    - rebuild-lab
  packages:
    - "Internal.*"
    - "Sector.*"

No CI default profile should run rebuilds.

Nix/dev environment

Prefer multiple dev shells/packages instead of a monster shell:

devShells.supply-light   # Python + dotnet + jq + git
devShells.supply-full    # supply-light + OSV/Syft/Grype/Scorecard if packaged
devShells.rebuild-lab    # heavier source checkout/build experiments

Nix pins Linux-native tooling. It must not pretend to solve Windows/DevExpress 12.2 build constraints.

Non-goals

  • No replacement for NuGet Audit, OSV-Scanner, Syft, Grype, Scorecard, SLSA, Sigstore or GitHub attestations.
  • No universal package trust verdict.
  • No broad ecosystem rebuild.
  • No required 007 gate.
  • No learned risk model.
  • No GUAC-like graph database in this issue.

Required controls

  1. Each adapter can be disabled without affecting others.
  2. Missing adapter/tool is reported as coverage limitation, not a clean result.
  3. Scorecard/posture evidence cannot prove package source correspondence.
  4. Attestation evidence cannot prove byte-for-byte reproducibility unless linked to rebuild evidence.
  5. Rebuild failures are classified: environment failure, restore failure, build mismatch, normalized binary mismatch, nupkg content mismatch.
  6. Rebuild scope is explicit and selected; no accidental broad sweep.
  7. No adapter performs policy decisions directly; policy evaluation remains separate.

Acceptance

  • Adapter design note exists for known-vuln, sbom, repo-posture, attestations, and rebuild-lab.
  • Profiles keep later slices disabled by default.
  • At least one external-tool adapter is fixture-tested without network access.
  • rebuild-lab has a dry-run/plan mode before any real build execution.
  • Reports clearly separate integrity, provenance, repository posture, source correspondence and reproducibility.
  • No issue in this track blocks the first SourceLink/PDB + NuGet policy slice.

Activity

  1. PhysShell commented on Nov 21, 2026

    @PhysShell
    OwnerAuthor

    Research reference (21.11.2026): Atomdrift Scan malicious-capability evidence

    Detailed Page · Scan · cleave diff · integration & exit contract

    Place only as a candidate external adapter in the deferred supply-evidence track. Its output concerns static indicators of malicious behavior in package/release bytes (including NuGet, Go, PE/DLL, scripts, VSIX, CI actions), not known-vulnerability evidence, SBOM, SourceLink/PDB provenance, SLSA attestation or reproducible build verification. A separate ML verdict must not silently become the OwnAudit authority to block or approve.

    Transferable contracts:

    • Preserve artifact identity: SHA-256, source/purl, scanner build, model/rules version, scan configuration, time, scope, findings and whether full or hash/corpus lookup supplied the result.
    • Preserve allow / block / unanalyzed / unavailable, plus scanner exit 0 / 1 / 2 / 3 / 4 (4 means incomplete rule set, never clean). Missing rules, skipped archive members or unsupported formats must be explicit coverage limitations, not successful absence-of-risk evidence.
    • Consider cleave diff as a narrow signal of changed potential capabilities from a previous pinned package revision. Capability != proven execution path and suspicious != malicious.
    • The upstream benchmark is author-run and domain-dependent: azoth reports PE recall@L25 82.78%, Go 70.56%, C# 39.95% on its own held-out corpus. Family-based generalization and our deployment accuracy remain UNKNOWN.

    Only after the first slice: a bounded offline fixture comparison against current OwnAudit baseline, GuardDog and SafeDep on identical supported inputs. Record incremental verified findings, false positives, unknowns, processing cost and value of version-to-version diff. GO-NARROW only for independently demonstrated value; otherwise retain evidence/status ideas without runtime adoption.

    This comment does not expand the issue's existing deferred sequence or authorize installation, vendor service, CI policy, real client artifacts, malware execution, uploads or implementation. 007 #156 owns preflight execution-risk inventory.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions