diff --git a/.github/workflows/fips-image.yml b/.github/workflows/fips-image.yml
index d096a12..a606c74 100644
--- a/.github/workflows/fips-image.yml
+++ b/.github/workflows/fips-image.yml
@@ -17,12 +17,19 @@ on:
       - .github/workflows/fips-image.yml
   workflow_dispatch:
 
+# `packages: write` moved from here to the job below on 2026-09-23 (OpenSSF Scorecard
+# Token-Permissions). A write granted at the top level is granted to every job the workflow will
+# ever have, including ones nobody has written yet; granted at the job level it is granted to the
+# one job that pushes the image, and a reviewer sees it next to the push.
 permissions:
   contents: read
-  packages: write
 
 jobs:
   image:
+    permissions:
+      contents: read
+      # The registry push at the end of this job, and nothing else here.
+      packages: write
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4.4.0
diff --git a/.github/workflows/gate.yml b/.github/workflows/gate.yml
index d119fd0..ef836a0 100644
--- a/.github/workflows/gate.yml
+++ b/.github/workflows/gate.yml
@@ -6,6 +6,14 @@ on:
   push:
   pull_request:
 
+# Least privilege for the workflow's token, added 2026-09-23 after OpenSSF Scorecard scored this
+# repository 0 on Token-Permissions: neither this workflow nor `package.yml` declared any
+# top-level permissions, so both ran with whatever the repository default happened to be. Every
+# job here reads the tree and writes nothing back, so the whole workflow gets read and any job
+# that later needs more has to say so itself and be seen saying it in review.
+permissions:
+  contents: read
+
 jobs:
   gate:
     runs-on: ubuntu-latest
@@ -108,6 +116,17 @@ jobs:
   # files that define it; this job never builds it.
   fips-tag:
     runs-on: ubuntu-latest
+    # `packages: read` is not decoration. The workflow-level block above narrows every job's token
+    # to what it names, and before that block existed these two jobs inherited the repository
+    # default, which grants read on every scope. Declaring `contents: read` at the top therefore
+    # took `packages: read` away, and this job's `docker login` to the registry stopped being able
+    # to see the FIPS image: it polled for the full thirty minutes and failed with "was not
+    # published", while the fips-image workflow had already reported the tag as present. Two
+    # checks disagreeing about whether an image exists is what an authorization failure looks like
+    # from the outside. Measured on this branch before it was merged.
+    permissions:
+      contents: read
+      packages: read
     outputs:
       ref: ${{ steps.tag.outputs.ref }}
     steps:
@@ -149,6 +168,11 @@ jobs:
   fips:
     needs: fips-tag
     runs-on: ubuntu-latest
+    # The runner pulls this container before the job's first step, using the credentials below,
+    # so without `packages: read` the job fails with `manifest unknown` and no output at all.
+    permissions:
+      contents: read
+      packages: read
     container:
       image: ${{ needs.fips-tag.outputs.ref }}
       credentials:
diff --git a/.github/workflows/package.yml b/.github/workflows/package.yml
index 5f98df8..84e4a79 100644
--- a/.github/workflows/package.yml
+++ b/.github/workflows/package.yml
@@ -40,6 +40,12 @@ on:
       - .github/workflows/package.yml
   workflow_dispatch:
 
+# Least privilege for the workflow's token (2026-09-23, OpenSSF Scorecard Token-Permissions). The
+# default is read; the one job that needs more, the attesting one, declares it at the job level
+# where a reviewer sees it beside the step that uses it.
+permissions:
+  contents: read
+
 jobs:
   # Slice 002. The bill of materials is derived from `mix.lock` alone: it is the same document
   # whichever runner produces it, and `mix trinity.sbom` needs the project compiled in the dev
diff --git a/test/workflows_test.exs b/test/workflows_test.exs
index 9393c85..65b8206 100644
--- a/test/workflows_test.exs
+++ b/test/workflows_test.exs
@@ -4,20 +4,31 @@ defmodule WorkflowsTest do
   @moduledoc """
   The continuous integration definitions are code, and they are checked like code.
 
-  Two properties. Every workflow and the Dependabot configuration parse as YAML, so a malformed
+  Three properties. Every workflow and the Dependabot configuration parse as YAML, so a malformed
   one fails here rather than silently not running: GitHub reports a broken workflow file on its
-  own page and not on the pull request, which is the wrong place for it to be discovered. And
-  every third-party action is pinned to a full commit hash rather than a tag, because a tag can
-  be moved to point at different code after review; this is the OpenSSF Scorecard
-  `Pinned-Dependencies` check, held here rather than waited for.
+  own page and not on the pull request, which is the wrong place for it to be discovered. Every
+  third-party action is pinned to a full commit hash rather than a tag, because a tag can be moved
+  to point at different code after review; this is the OpenSSF Scorecard `Pinned-Dependencies`
+  check, held here rather than waited for. And every workflow declares a top-level `permissions:`
+  block granting no more than read, so that a job wanting to write has to say so itself: Scorecard's
+  `Token-Permissions` check, which this repository scored **zero** on until 2026-09-23 because two
+  workflows declared nothing at all and ran with whatever the repository default happened to be.
+
+  The third property is the one worth a sentence about why a test rather than a habit. A top-level
+  write is granted to every job the workflow will ever have, including the ones nobody has written
+  yet, and the cost of forgetting is invisible: nothing fails, the token is simply broader than the
+  work needs. That is exactly the class of thing a build should hold rather than a reviewer.
 
   The population comes from the tree (`git ls-files .github`), so a workflow added without a
-  thought for either property is caught by its own existence.
+  thought for any of the three is caught by its own existence.
   """
   use ExUnit.Case, async: true
 
   @action_ref ~r{uses:\s*([A-Za-z0-9_.-]+/[A-Za-z0-9_./-]+)@(\S+)}
 
+  # Anything that is not one of these is a write, whatever it is called.
+  @read_only ~w(read none)
+
   defp yaml_files do
     {out, 0} = System.cmd("git", ["ls-files", ".github"])
 
@@ -54,4 +65,38 @@ defmodule WorkflowsTest do
                "tag can be moved to point at different code after it was reviewed"
     end
   end
+
+  test "every workflow declares a top-level permissions block, and it grants no write" do
+    for file <- yaml_files(), String.contains?(file, "/workflows/") do
+      permissions = file |> read_yaml() |> Map.get("permissions")
+
+      assert permissions != nil,
+             "#{file} declares no top-level permissions, so its token is whatever the " <>
+               "repository default happens to be. Declare `permissions: {contents: read}` and " <>
+               "let any job that needs to write say so itself"
+
+      for {scope, level} <- top_level_scopes(permissions) do
+        assert level in @read_only,
+               "#{file} grants `#{scope}: #{level}` at the top level, which grants it to every " <>
+                 "job the workflow will ever have. Move it to the job that needs it"
+      end
+    end
+  end
+
+  # `permissions:` is either a map of scopes, or one of the shorthands `read-all`/`write-all`.
+  defp top_level_scopes("read-all"), do: []
+  defp top_level_scopes("write-all"), do: [{"(every scope)", "write"}]
+  defp top_level_scopes(map) when is_map(map), do: map
+  defp top_level_scopes(other), do: [{"(unrecognised)", inspect(other)}]
+
+  defp read_yaml(file) do
+    {json, 0} =
+      System.cmd("python3", [
+        "-c",
+        "import sys,yaml,json; json.dump(yaml.safe_load(open(sys.argv[1])), sys.stdout)",
+        file
+      ])
+
+    Jason.decode!(json)
+  end
 end