Binskim Analysis

BinSkim Official Results Analysis

Investigate and interpret BinSkim findings from official 1ES SDL pipelines. This skill helps you understand what the central portal reports, what gets filtered, and how to determine fix ownership.

For running BinSkim locally against a repo, use the binskim-scan skill instead.

When to Use This Skill

• Investigating BinSkim findings reported in the central SDL portal
• Understanding why the portal shows (or doesn't show) specific findings
• Comparing local scan results with official pipeline results
• Determining whether a finding is fixable in the repo or requires an upstream vendor fix
• Questions like "why does the portal show BA2008 on X.dll?", "what rules apply to my repo?"

How Official SDL Scanning Works

Official pipelines run BinSkim via Guardian, not directly. Results go through filtering before reaching the portal:

BinSkim runs -> raw binskim.sarif (ALL findings)
     |
Guardian merges/filters -> Results.sarif (subset)
     |
Results.sarif uploaded -> Central portal

Your local scan will find more issues than the portal reports. This is expected — the portal only shows findings for rules with SDL policy requirement mappings for your org. See references/guardian-filtering.md for the full filtering pipeline.

Downloading SDL Artifacts

To investigate official results, download the SDL artifact from the repo's official build:

# Use AzDO tools or ado-dnceng-pipelines_download_artifact
# Artifact naming pattern:
#   Most repos:  drop_build_Windows_x64_sdl_analysis
#   VMR repos:   drop_VMR_Vertical_Build_Windows_x64_sdl_analysis
# Some repos have multiple SDL legs — check ALL OS legs (Windows, Linux, macOS)
# Each OS build produces different native binaries with potentially different findings
#
# ⚠️ NON-WINDOWS SDL ARTIFACTS MAY BE EMPTY. Some pipelines may have a bug where
#    BinSkim only runs on Windows SDL legs. Linux/macOS SDL artifacts may exist but contain 0-byte
#    SARIF files. Always check file sizes before parsing — don't assume non-Windows legs
#    have results.
#
# ⚠️ LARGE PIPELINE ARTIFACTS: For large artifacts (>100MB), the AzDO REST API may return
#    an HTML sign-in page if you use $format=zip directly. Instead:
#    1. GET the artifact metadata first:
#       GET .../_apis/build/builds/{buildId}/artifacts?artifactName={name}&api-version=7.1
#    2. Extract the downloadUrl from the response: .resource.downloadUrl
#    3. GET that URL with a bearer token to download the actual content
#    This CDN-based download works reliably for any artifact size.

# Key files inside the artifact:
# - binskim/001/binskim.sarif    <- Raw BinSkim output (everything found)
# - Results.sarif                 <- Guardian-merged output (portal sees this)
# - break/001/options.json        <- Break policy (which tools can fail the build)
# - .gdnbaselines                 <- Auto-generated baseline suppressions

Downloading via MCP when the agent host suppresses binary tool output

Some agent CLI hosts run the MCP pipelines_artifact download action successfully but silently drop the embedded-resource zip (you see "Tool ran without output or errors" and no file lands anywhere). When that happens, call the MCP HTTP endpoint directly from PowerShell and decode the base64 blob yourself:

1. Find the MCP servers config (the file path is in env, but typically): Get-ChildItem $env:TEMP -Filter "copilot-mcp-*.json" | Sort LastWriteTime -Descending | Select -First 1 | Get-Content -Raw — locate the ado-dnceng (or ado-dnceng-public) entry's url and capture it in a variable (the port is assigned per session, so it changes every time):

   $mcpUrl = (Get-Content (Get-ChildItem $env:TEMP -Filter 'copilot-mcp-*.json' | Sort LastWriteTime -Descending | Select -First 1) -Raw | ConvertFrom-Json).mcpServers.'ado-dnceng'.url
   

2. POST a tools/call JSON-RPC request (uses $mcpUrl from step 1 — do not hardcode the port):

   $req = @{ jsonrpc='2.0'; id=2; method='tools/call'; params=@{
     name='pipelines_artifact'
     arguments=@{ action='download'; buildId=<id>; project='internal'; artifactName='<name>' }
   } } | ConvertTo-Json -Compress -Depth 10
   $msg = [System.Net.Http.HttpRequestMessage]::new('POST',$mcpUrl)
   $msg.Headers.Accept.ParseAdd('application/json, text/event-stream')
   $msg.Content = [System.Net.Http.StringContent]::new($req,[Text.Encoding]::UTF8,'application/json')
   $client = [System.Net.Http.HttpClient]::new(); $client.Timeout = [TimeSpan]::FromMinutes(15)
   $body = $client.SendAsync($msg).Result.Content.ReadAsStringAsync().Result
   

3. Response is Server-Sent Events: find the data: line and ConvertFrom-Json. The zip is at result.content[0].resource.blob (base64). Write with [IO.File]::WriteAllBytes($path, [Convert]::FromBase64String($blob)) then Expand-Archive.
4. Reusable script template: see C:\temp\Download-AdoArtifact.ps1 if previously generated, or recreate from the snippet above.

Notes:

• The MCP-served zip is dedup-compressed and is much smaller than the artifactsize reported by the artifact list call (an 86 MB "artifactsize" Windows_x64 SDL artifact downloads as a ~5–6 MB zip but contains the full 40+ MB raw and merged SARIF after extraction). Don't be alarmed by the size delta.
• az login / Connect-AzAccount may hang indefinitely in non-interactive agent shells — don't rely on them. The MCP HTTP path above uses the MCP server's existing auth so no extra login is needed.

Large SARIF files: For repos with many binaries (especially the VMR), raw and merged SARIF can be 50-80MB+. Use Get-Content -Raw | ConvertFrom-Json and stream results with Group-Object.

SDL does NOT run on PR validation builds. It only runs on official/CI pipelines (gated by Build.Reason != PullRequest). Users must manually queue the official pipeline for SDL results before merging.

Per-repo pipeline names — see references/repo-profiles-index.md for the index of per-repo profile files (one per repo). Load only the file for the repo you're investigating.

SDL Artifact Structure

The SDL artifact contains a rich directory structure. Key files and what they tell you:

<SDL artifact>/
  binskim/
    001/binskim.sarif         # Raw BinSkim output for scan directory 1 (e.g., bin/)
    001/.gdntoolinput         # Explicit file list — exactly what was scanned
    002/binskim.sarif         # Raw BinSkim output for scan directory 2 (e.g., symbols/)
    002/.gdntoolinput         # Explicit file list for directory 2
  prefast/                    # Native C++ static analysis (separate from BinSkim)
  roslynanalyzers/            # Managed code analysis
  break/001/options.json      # Break policy — which tools can fail the build
  Results.sarif               # Guardian-merged & deduplicated (this goes to portal)
  Results3.sarif              # Cumulative raw output from ALL tool phases
  .gdnbaselines               # Baselined finding signatures (managed by Guardian portal)
  .gdnsuppress                # Suppressed findings (same set, filtered from S360)

Multiple phases: Most tools run in 4 phases (001-004); BinSkim typically has 2 (one per staging directory). Results3.sarif aggregates all phases — expect hundreds of "runs" (e.g., 417 = 310 roslynanalyzers + 100 prefast + 7 binskim). Results.sarif deduplicates these down (e.g., to 42 unique runs).

.gdntoolinput is the source of truth for what was actually scanned. It's an explicit file list (one quoted path per line), NOT a recursive glob. If a binary isn't in this file, BinSkim didn't see it.

.gdnbaselines are NOT repo-committed — they're auto-generated pipeline artifacts used for break-on-new-only tracking. .gdnsuppress files CAN be checked into source to suppress specific findings, but this is rarely done in dotnet repos. Neither can weaken SDL policy — the tenant/org policy sets the floor.

Comparing Raw vs Merged SARIF

Always compare both files to understand what Guardian filtered:

$raw = (Get-Content "binskim\001\binskim.sarif" -Raw | ConvertFrom-Json).runs[0].results
# IMPORTANT: Results.sarif contains multiple runs (BinSkim, roslynanalyzers, prefast, etc.)
# Filter to the BinSkim run — don't assume runs[0] is BinSkim
$mergedSarif = Get-Content "Results.sarif" -Raw | ConvertFrom-Json
$binskimRun = $mergedSarif.runs | Where-Object { $_.tool.driver.name -like '*BinSkim*' } | Select-Object -First 1
$merged = if ($binskimRun) { $binskimRun.results } else { $mergedSarif.runs[0].results }

Write-Host "Raw BinSkim findings: $($raw.Count)"
Write-Host "After Guardian filtering: $($merged.Count)"
Write-Host "`nRaw findings by rule:"
$raw | Group-Object ruleId | ForEach-Object {
    $levels = ($_.Group | ForEach-Object { $_.level } | Sort-Object -Unique) -join ","
    Write-Host "  $($_.Name): count=$($_.Count) levels=[$levels]"
}

The delta between raw and merged is what Guardian filtered out based on SDL policy. See references/guardian-filtering.md for why things get filtered and the suppression mechanisms available.

Service Tree and Org Awareness

Which rules the portal reports depends on es-metadata.yml — the routing.defaultAreaPath.org field determines your SDL policy scope. Different orgs surface different rules. For example, the devdiv org (most dotnet/* repos) treats BA2008/BA2009/BA2021 as SDL-required (10203-mapped), while the nettel org surfaces BA2004/BA2027 as informational findings in the portal. Neither BA2004 nor BA2027 is SDL-required at the company level (no 10203 mapping) — but informational portal findings still often get triaged like requirements, so they are worth fixing. See references/repo-perfview.md for a worked example.

See references/binskim-rules.md for the full rules tables and the observed portal requirements by org.

Coverage Gap Detection

After examining results, check whether the scan covers everything the repo ships:

1. List all published artifacts from the pipeline YAML
2. Compare with .gdntoolinput file lists in the SDL artifact — these are the explicit files Guardian scanned. Not a recursive directory walk.
3. Report gaps: "This repo ships X.dll (native C++ binary) but it's not in the .gdntoolinput scan targets — likely because it isn't staged to the bin/ or symbols/ directories that Guardian scans."

Common gap — unstaged binaries: Native C++ binaries built by .vcxproj often output to project-local Release\ directories, not the main artifacts\bin\ staging area. Guardian generates .gdntoolinput from the staging directories, so these binaries get Prefast coverage (source analysis) but miss BinSkim (binary analysis). This is a pipeline artifact staging config issue.

Common gap — missing OS coverage: Each OS build configuration produces different native binaries (e.g., Linux .so, macOS .dylib, Windows .dll). If the official pipeline only runs BinSkim on Windows SDL legs, Linux and macOS native binaries are not scanned at all. Flag this clearly as a coverage gap that needs fixing. Note: this gap is typically caused by the 1ES/Guardian SDL template using Windows-centric glob patterns (matching .dll/.exe/.sys) rather than a BinSkim limitation — BinSkim can scan PE, ELF, and Mach-O binaries on any OS using magic-byte detection. To close the gap, either configure broader glob patterns in the existing SDL legs, or add BinSkim scanning steps that target non-Windows artifacts with appropriate patterns (*.dylib, *.so, or **).

Classification Framework

When analyzing findings, classify each into one of four buckets to determine the right action:

To classify: check the SARIF artifact path for Shipping vs NonShipping, and search the repo for the binary's source project (.vcxproj/.csproj) vs NuGet package origin.

Fix Ownership

Not all findings can be fixed in the repo where they're reported:

VMR (dotnet/dotnet): Findings map to source sub-repos. Look at the artifact path in the SARIF (e.g., src/arcade/artifacts/... means fix goes to dotnet/arcade).

Common misconception

<ControlFlowGuard>Guard</ControlFlowGuard> in a .csproj does nothing. The C# compiler has no /guard:cf support. This MSBuild property only works in .vcxproj (MSVC C++). BA2008 only fires on native PE binaries.

Non-Arcade Repos

Some repos (e.g., microsoft/perfview) don't use arcade infrastructure:

• No eng/common/sdl/ — BinSkim config is in the 1ES template or pipeline YAML directly
• Build output may be in src\bin\ instead of artifacts\
• Look for .ado.yml or .pipelines/ for pipeline config
• The general analysis approach still works — find the SDL artifact, compare raw vs merged SARIF

Anti-Patterns

Don't trust the central portal as the complete picture. Download the raw binskim.sarif and compare with Results.sarif. The portal shows only findings for rules with SDL policy mappings.

Don't assume "not in the portal" means "not a problem." Guardian filtering removes findings without SDL policy mappings for your org. These may still be good security practice.

Don't confuse break policy with reporting. Break policy controls whether BinSkim fails the build. It does NOT control what gets reported to the portal. Findings can be reported without breaking the build.

Don't assume Windows-only scanning is sufficient. If the repo ships native binaries for Linux or macOS (.so, .dylib), those binaries may have different BinSkim findings than their Windows counterparts. If official SDL legs only cover Windows, flag this as a gap — the pipeline should be updated to run BinSkim on all OS configurations that produce shipped native artifacts.

Comparing Local Scan vs Official Run

When the user has run the binskim-scan skill locally and wants to validate against official results:

Key reconciliation questions:

1. Do all local errors also appear in official? If not, they may be in test/NuGet paths that official doesn't scan.
2. Do any official findings NOT appear locally? Could mean the binary wasn't built locally (e.g., missing C++ toolset).
3. Are there binaries in the official .gdntoolinput that the local scan missed? Unlikely with recursive scan — the other direction (local finds more) is expected.
4. Does official scan native C++ binaries that ship? Check .gdntoolinput — this is a known gap for repos where C++ output goes to non-standard directories.

Don't report NuGet transitive dependency findings as repo issues. If libSkiaSharp.dll fails BA2008, that's SkiaSharp's issue — unless this repo ships it.

References

• BinSkim rules (all platforms): references/binskim-rules.md
• Guardian filtering deep-dive: references/guardian-filtering.md
• Per-repo profiles: references/repo-profiles-index.md (index — one .md per repo)
• Arcade SDL infrastructure: references/arcade-sdl.md
• Local scanning: Use the binskim-scan skill