diff --git a/.github/workflows/sdk-regression.yml b/.github/workflows/sdk-regression.yml
index 35f60b754..ad78d7ce9 100644
--- a/.github/workflows/sdk-regression.yml
+++ b/.github/workflows/sdk-regression.yml
@@ -21,13 +21,16 @@ jobs:
   regression:
     name: regression
     runs-on: ubuntu-latest
-    # Least-privilege: read code/PR data and write commit statuses only. Note an
-    # author-permission guard below (check-access) already restricts triggering
-    # to write/admin collaborators (CWE-284 / PER-8610).
+    # Least-privilege GITHUB_TOKEN: read code/PR data, write commit statuses, and
+    # read this run's Actions job metadata (for the job log URL) only. Triggering
+    # is further gated by the author-permission check (check-access) below, which
+    # restricts it to write/admin collaborators. The powerful cross-repo dispatch
+    # PAT is NOT granted here; it is injected only into the single trigger step.
     permissions:
       contents: read
       pull-requests: read
       statuses: write
+      actions: read
     # Run when either:
     #  - a maintainer comments RUN_REGRESSION on a PR, or
     #  - the workflow is dispatched manually (CLI branch + optional per-SDK refs).
@@ -140,17 +143,22 @@ jobs:
         # failure must never gate the regression itself.
         continue-on-error: true
         with:
-          github_token: ${{ secrets.WORKFLOW_DISPATCH_ACTIONS_TOKEN }}
+          # Third-party action: only reads this run's job metadata on percy/cli,
+          # so the built-in GITHUB_TOKEN (actions: read) is sufficient. The
+          # cross-repo dispatch PAT is deliberately not exposed to it.
+          github_token: ${{ github.token }}
           job_name: "regression (${{ matrix.repo }})"
           # The fan-out matrix is >30 jobs (currently 35); the action defaults to
           # per_page=30, so jobs on page 2 resolve to null and exit 1. Cover the
           # whole matrix (GitHub jobs API max page size is 100).
           per_page: 100
       - name: Output Current Job Log URL
-        run: echo ${{ steps.jobs.outputs.html_url }}
+        run: echo ${{ steps.job-url.outputs.html_url }}
       - uses: actions/github-script@f891eff65186019cbb3f7190c4590bc0a1b76fbc # v4.1.0
         with:
-          github-token: ${{ secrets.WORKFLOW_DISPATCH_ACTIONS_TOKEN }}
+          # Writes a commit status on percy/cli itself (job-level statuses: write),
+          # so the built-in GITHUB_TOKEN is used instead of the cross-repo PAT.
+          github-token: ${{ github.token }}
           script: |
             const { owner, repo } = context.repo;
             const sha = '${{ steps.pr.outputs.head_sha }}'
@@ -201,6 +209,9 @@ jobs:
         with:
           owner: percy
           repo: ${{ steps.split.outputs._0 }}
+          # Only step that strictly needs the cross-repo dispatch PAT: it triggers
+          # test.yml in a *different* percy repo, which the built-in GITHUB_TOKEN
+          # cannot do. Injection is scoped to this step alone.
           github_token: ${{ secrets.WORKFLOW_DISPATCH_ACTIONS_TOKEN }}
           # Most SDKs expose `test.yml`; a few use a differently-named workflow
           # that carries the @percy/cli inject step:
@@ -214,7 +225,9 @@ jobs:
       - name: Update Status
         uses: actions/github-script@f891eff65186019cbb3f7190c4590bc0a1b76fbc # v4.1.0
         with:
-          github-token: ${{ secrets.WORKFLOW_DISPATCH_ACTIONS_TOKEN }}
+          # Writes a commit status on percy/cli itself (job-level statuses: write),
+          # so the built-in GITHUB_TOKEN is used instead of the cross-repo PAT.
+          github-token: ${{ github.token }}
           script: |
             const { owner, repo } = context.repo;
             const sha = '${{ steps.pr.outputs.head_sha }}'