diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index ac86d8858..d3ce4751b 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -391,11 +391,17 @@ jobs:
     name: Playwright E2E (production)
     runs-on: ubuntu-latest
     # A backstop above the test step's own cap, not the primary bound: a job
-    # timeout is a cancellation and would skip the rollback below, whereas the
-    # step cap marks the step failed and keeps it reachable.
+    # timeout cancels the test step, and a cancelled step is not a verdict the
+    # rollback job acts on, whereas the step cap marks it failed.
     timeout-minutes: 30
     needs: [production-smoke-tests]
     if: github.ref == 'refs/heads/main' && github.event_name == 'push'
+    # The test step's own outcome, for rollback-production-on-e2e-failure. Not
+    # the job's result: a failed checkout, npm ci or Chrome report fails the job
+    # too, and none of them says anything about whether production is healthy.
+    # Only tests that ran and failed roll production back.
+    outputs:
+      tests: ${{ steps.e2e.outcome }}
     # The browser is the Google Chrome the runner image ships
     # (e2e/playwright.config.ts) rather than one installed here: `npx
     # playwright install --with-deps` puts an apt update and Playwright's CDN in
@@ -418,9 +424,9 @@ jobs:
       - name: Report the runner's Chrome
         run: node e2e/report-chrome.js
       # Bounded at the step rather than relying on the job cap: a step timeout
-      # marks the step failed, which still trips the rollback below, whereas a
-      # job timeout is a cancellation and would skip it. A production wedged
-      # badly enough to hang Playwright is exactly when the rollback is wanted.
+      # marks the step failed, which still trips the rollback job, whereas a
+      # job timeout cancels it and would not. A production wedged badly enough
+      # to hang Playwright is exactly when the rollback is wanted.
       - name: Run E2E tests (production)
         id: e2e
         # See the staging job: the step cap has to clear nodes.spec.ts's retried
@@ -430,13 +436,29 @@ jobs:
         env:
           CI: "true"
           RADAR_API_KEY: ${{ secrets.RADAR_API_KEY }}
-      # Scoped to the test step's own outcome, as the production smoke job is.
-      # A bare failure() also fires when checkout, npm ci or the Chrome report
-      # fails, and anything that breaks before the tests run is a statement
-      # about the runner, not about production. Those leave the job red for a
-      # human to read; only tests that ran and failed roll production back.
-      - name: Rollback production on E2E failure
-        if: failure() && steps.e2e.outcome == 'failure'
+
+  # A job of its own, so that production's root key is never handed to a job
+  # that has just run npm's install scripts. test_deploy_workflows.py holds
+  # every workflow to that.
+  rollback-production-on-e2e-failure:
+    name: Roll back production if its E2E tests failed
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    needs: [e2e-prod]
+    # See rollback-production-on-deploy-failure: a following push's deploy must
+    # not be inside the deploy directory while rollback.sh is.
+    concurrency:
+      group: production-deploy
+      cancel-in-progress: false
+    # !cancelled() rather than always(): e2e-prod failing is when this has to
+    # run, but an operator cancelling the run, say over a known flake, must
+    # still be able to stop it. The main-push gate is repeated so no other event
+    # can open production SSH.
+    if: >-
+      !cancelled() && github.ref == 'refs/heads/main' && github.event_name == 'push' &&
+      needs.e2e-prod.outputs.tests == 'failure'
+    steps:
+      - name: Roll back production
         uses: appleboy/ssh-action@0ff4204d59e8e51228ff73bce53f80d53301dee2 # v1.2.5
         with:
           host: ${{ secrets.SERVER_HOST }}
@@ -934,8 +956,8 @@ jobs:
 
             # Reached only once the app answered /api/health. The box is no
             # longer mid-deploy, so drop the marker; any failure after this
-            # point is a smoke/E2E failure, which the rollback steps on those
-            # jobs already cover.
+            # point is a smoke/E2E failure, which the smoke job's rollback step
+            # and rollback-production-on-e2e-failure already cover.
             rm -f ${{ env.APP_DIR }}/.deploy-in-progress
 
   rollback-production-on-deploy-failure:
@@ -959,9 +981,9 @@ jobs:
     # would then SSH into production to consider rolling back a deploy that
     # never ran. Keyed to the deploy job's own result, it cannot.
     #
-    # This is the gap the two existing rollbacks do not cover. Both hang off
-    # jobs DOWNSTREAM of the deploy (production-smoke-tests, e2e-prod), and a
-    # failed deploy skips those jobs and their rollback steps with them.
+    # This is the gap the two other rollbacks do not cover. Both hang off jobs
+    # DOWNSTREAM of the deploy (production-smoke-tests, e2e-prod), and a failed
+    # deploy skips those jobs, and so the rollbacks keyed to them.
     # 'cancelled' as well as 'failure'. Cancelling mid-deploy leaves production
     # exactly as half-deployed as a crash does, and deploy-production sets
     # cancel-in-progress: false, so this can only be a human pressing cancel —
@@ -1368,7 +1390,7 @@ jobs:
           echo ""
           smoke_summary
           [ "$FAIL" -eq 0 ] || exit 1
-      # Auto-rollback on smoke failure, mirroring the prod E2E job. A failed
+      # Auto-rollback on smoke failure, as on an E2E failure. A failed
       # smoke test means a bad build is live serving prod; revert to
       # last-known-good rather than leaving it up for a human to catch. The
       # condition is scoped to the smoke step's own outcome, so a failure in the
@@ -1383,7 +1405,7 @@ jobs:
           key: ${{ secrets.SERVER_SSH_KEY }}
           envs: APP_DIR
           script: |
-            # See the identical assertion in e2e-prod's rollback step.
+            # See the identical assertion in rollback-production-on-e2e-failure.
             [ "$(hostname)" = "retina-prod" ] || { echo "::error::Refusing to roll back on '$(hostname)' — expected retina-prod."; exit 1; }
             echo "Production smoke tests failed — rolling back production..."
             cd ${{ env.APP_DIR }} && bash deploy/rollback.sh
diff --git a/backend/tests/test_deploy_workflows.py b/backend/tests/test_deploy_workflows.py
index bae525009..2ea43cfd2 100644
--- a/backend/tests/test_deploy_workflows.py
+++ b/backend/tests/test_deploy_workflows.py
@@ -4,6 +4,7 @@
 and pin what must hold in every copy.
 """
 
+import functools
 import re
 from pathlib import Path
 
@@ -12,6 +13,12 @@
 
 WORKFLOWS = Path(__file__).resolve().parents[2] / ".github" / "workflows"
 
+
+@functools.cache
+def _workflow(workflow: str) -> dict:
+    return yaml.safe_load((WORKFLOWS / workflow).read_text())
+
+
 DEPLOYS = [
     pytest.param("ci.yml", "deploy-production", id="production"),
     pytest.param("staging-deploy-verify.yml", "deploy", id="staging"),
@@ -26,7 +33,7 @@
 
 
 def _job(workflow: str, job: str) -> dict:
-    jobs = yaml.safe_load((WORKFLOWS / workflow).read_text())["jobs"]
+    jobs = _workflow(workflow)["jobs"]
     assert job in jobs, f"{workflow} has no job {job!r}"
     return jobs[job]
 
@@ -205,22 +212,165 @@ def test_rollback_job_reads_back_what_the_deploy_wrote(workflow, deploy, rollbac
     assert "always()" in condition
 
 
+def _allowed(workflow: str, job: str, values: dict[str, str], *, cancelled: bool = False) -> bool:
+    """A job's `if:`, evaluated with each context field replaced by a value."""
+    condition = " ".join(_job(workflow, job)["if"].split())
+    for field, value in values.items():
+        # Whole fields only: github.ref must not also rewrite github.ref_name.
+        condition = re.sub(rf"(?<![\w.-]){re.escape(field)}(?![\w-])", repr(value), condition)
+    # The rest is equality and boolean operators over concrete strings, whose
+    # semantics are the same in Actions and Python. Any other function is left
+    # undefined, and so fails the test rather than guessing.
+    condition = condition.replace("!cancelled()", repr(not cancelled)).replace("always()", "True")
+    assert not re.search(r"!(?!=)", condition), f"no Python equivalent for a negation in {condition!r}"
+    condition = condition.replace("&&", " and ").replace("||", " or ")
+    return eval(condition, {"__builtins__": {}})
+
+
 @pytest.mark.parametrize("ref", ["refs/heads/main", "refs/heads/topic", "refs/pull/417/merge"])
 @pytest.mark.parametrize("event", ["push", "pull_request", "workflow_dispatch"])
 @pytest.mark.parametrize("result", ["failure", "cancelled", "success", "skipped"])
-def test_production_rollback_requires_a_failed_or_cancelled_main_push(ref, event, result):
+@pytest.mark.parametrize("cancelled", [False, True])
+def test_production_rollback_requires_a_failed_or_cancelled_main_push(ref, event, result, cancelled):
     # A cancelled PR run can report a skipped deploy as cancelled. Evaluate
     # the real job condition: the remote marker check is too late to prevent
     # an unauthorized workflow event from opening a production SSH session.
-    condition = " ".join(_job("ci.yml", "rollback-production-on-deploy-failure")["if"].split())
-    for field, value in (
-        ("github.ref", ref),
-        ("github.event_name", event),
-        ("needs.deploy-production.result", result),
-    ):
-        condition = condition.replace(field, repr(value))
-    # This expression uses only equality and boolean operators, whose
-    # semantics are the same for these concrete strings in Actions/Python.
-    condition = condition.replace("always()", "True").replace("&&", "and").replace("||", "or")
-    allowed = eval(condition, {"__builtins__": {}})
+    # Cancelling the run does not stop it: a cancel mid-deploy leaves the box
+    # as half-deployed as a crash does.
+    allowed = _allowed(
+        "ci.yml",
+        "rollback-production-on-deploy-failure",
+        {"github.ref": ref, "github.event_name": event, "needs.deploy-production.result": result},
+        cancelled=cancelled,
+    )
     assert allowed is (ref == "refs/heads/main" and event == "push" and result in {"failure", "cancelled"})
+
+
+# ── Third-party code never shares a job with a droplet key ───────────────────
+# A package manager runs install scripts from its registry, an action runs its
+# author's code, and a container image is someone else's filesystem. Any of them
+# can tamper with the environment a later step hands a root key to, so a job
+# that runs one must not be able to read a key. Rollbacks that follow such a job
+# run in a job of their own.
+
+# A tripwire for the honest mistake, not a sandbox: flags may sit between a
+# tool and its verb, and a script a step calls is not followed.
+THIRD_PARTY_CODE = re.compile(
+    r"\b(?:(?:npm|pip3?|apt(?:-get)?|gem|cargo|go|bun|poetry)\b[^\n;&|]*\binstall"
+    r"|npm\b[^\n;&|]*\b(?:ci|i|exec)|npx|yarn|pnpm|pipx|uvx|apk\s+add|curl\b[^\n;&]*\|\s*(?:ba)?sh"
+    r"|uv\b[^\n;&|]*\b(?:sync|add|run|pip|tool)|pre-commit"
+    r"|docker(?:-compose|\b[^\n;&|]*\b(?:run|build|pull|compose)))\b",
+    re.IGNORECASE,
+)
+# The actions a key-holding job may use: the checkout, and the ssh transport the
+# key is for. Any other action counts as third-party code.
+KEY_JOB_ACTIONS = ("actions/checkout@", "appleboy/ssh-action@")
+# The droplet keys are all named *SSH*KEY. A dynamic lookup, the whole secrets
+# object, or `secrets: inherit` could be any of them. Case-blind, as Actions
+# expressions are.
+SSH_KEY = re.compile(
+    r"\bsecrets(?:\.\w*SSH\w*KEY\w*\b|\[\s*['\"]\w*SSH\w*KEY\w*['\"]\s*\]|\[\s*(?!['\"])|:\s*inherit\b)"
+    r"|\btoJSON\(\s*secrets\s*\)",
+    re.IGNORECASE,
+)
+
+
+def _workflow_files() -> list[Path]:
+    return sorted([*WORKFLOWS.glob("*.yml"), *WORKFLOWS.glob("*.yaml")])
+
+
+def _workflow_jobs() -> list:
+    return [
+        pytest.param(path.name, name, job, id=f"{path.name}:{name}")
+        for path in _workflow_files()
+        for name, job in _workflow(path.name)["jobs"].items()
+    ]
+
+
+def _runs_third_party_code(job: dict) -> bool:
+    steps = job.get("steps", [])
+    return (
+        "container" in job
+        or "services" in job
+        # Another repository's reusable workflow.
+        or ("uses" in job and not str(job["uses"]).startswith("./"))
+        or any("uses" in step and not str(step["uses"]).startswith(KEY_JOB_ACTIONS) for step in steps)
+        or any(THIRD_PARTY_CODE.search(step.get("run", "")) for step in steps)
+    )
+
+
+def test_the_guard_recognises_what_the_jobs_run():
+    # Guards the guard: patterns that matched nothing would pass every job.
+    assert _runs_third_party_code(_job("ci.yml", "web-build"))
+    assert _runs_third_party_code(_job("ci.yml", "backend-tests"))
+    assert _runs_third_party_code(_job("ci.yml", "lint"))
+    assert _runs_third_party_code(_job("ci.yml", "e2e-prod"))
+    assert _runs_third_party_code({"services": {"db": {"image": "postgres"}}, "steps": []})
+    assert _runs_third_party_code({"steps": [{"uses": "docker://alpine:3"}]})
+    assert _runs_third_party_code({"steps": [{"uses": "./.github/actions/setup-uv"}]})
+    assert _runs_third_party_code({"steps": [{"run": "uv run python scripts/check.py"}]})
+    assert _runs_third_party_code({"steps": [{"run": "docker run --rm some/image"}]})
+    assert _runs_third_party_code({"steps": [{"run": "npm --prefix frontend ci"}]})
+    assert _runs_third_party_code({"steps": [{"run": "sudo apt-get -y install jq"}]})
+    assert _runs_third_party_code({"uses": "someorg/repo/.github/workflows/x.yml@main"})
+    assert not _runs_third_party_code(_job("ci.yml", "staging"))
+    assert not _runs_third_party_code(_job("ci.yml", "deploy-production"))
+    assert not _runs_third_party_code(_job("ci.yml", "production-smoke-tests"))
+    assert SSH_KEY.search(yaml.safe_dump(_job("ci.yml", "deploy-production")))
+    assert SSH_KEY.search("key: ${{ secrets['STAGING_SSH_KEY'] }}")
+    assert SSH_KEY.search("key: ${{ secrets[format('{0}_SSH_KEY', inputs.env)] }}")
+    assert SSH_KEY.search("ALL: ${{ toJson(secrets) }}")
+    assert SSH_KEY.search("key: ${{ secrets.server_ssh_key }}")
+    assert SSH_KEY.search("key: ${{ secrets.DEPLOY_SSH_PRIVATE_KEY }}")
+    assert SSH_KEY.search(yaml.safe_dump({"uses": "someorg/repo/.github/workflows/x.yml@main", "secrets": "inherit"}))
+    assert not SSH_KEY.search("RADAR_API_KEY: ${{ secrets.RADAR_API_KEY }}")
+
+
+@pytest.mark.parametrize(("workflow", "name", "job"), _workflow_jobs())
+def test_no_job_that_runs_third_party_code_can_read_a_droplet_key(workflow, name, job):
+    if _runs_third_party_code(job):
+        assert not SSH_KEY.search(yaml.safe_dump(job)), f"{workflow}:{name} runs third-party code"
+
+
+@pytest.mark.parametrize("path", _workflow_files(), ids=lambda path: path.name)
+def test_no_droplet_key_is_handed_to_every_job(path):
+    # Workflow-level env reaches every job, the npm ones included.
+    assert not SSH_KEY.search(yaml.safe_dump(_workflow(path.name).get("env", {})))
+
+
+# ── Production's E2E rollback ────────────────────────────────────────────────
+
+
+def test_the_e2e_rollback_acts_on_the_test_step_alone():
+    # A failed pull, checkout or npm ci fails e2e-prod as well, and says
+    # nothing about production. Only the test step's own verdict may count.
+    e2e = _job("ci.yml", "e2e-prod")
+    assert [step for step in e2e["steps"] if step.get("id") == "e2e" and "test:e2e:prod" in step["run"]]
+    assert e2e["outputs"]["tests"] == "${{ steps.e2e.outcome }}"
+
+
+def test_the_e2e_rollback_holds_the_production_lock_on_the_runner():
+    rollback = _job("ci.yml", "rollback-production-on-e2e-failure")
+    assert "container" not in rollback
+    assert rollback["needs"] == ["e2e-prod"]
+    assert rollback["concurrency"] == {"group": "production-deploy", "cancel-in-progress": False}
+    script = _script("ci.yml", "rollback-production-on-e2e-failure")
+    identity = _index(script, re.escape('[ "$(hostname)" = "retina-prod" ]'))
+    assert identity < _index(script, r"\bbash deploy/rollback\.sh$")
+
+
+@pytest.mark.parametrize("ref", ["refs/heads/main", "refs/heads/topic", "refs/pull/417/merge"])
+@pytest.mark.parametrize("event", ["push", "pull_request", "workflow_dispatch"])
+@pytest.mark.parametrize("tests", ["failure", "cancelled", "success", "skipped", ""])
+@pytest.mark.parametrize("cancelled", [False, True])
+def test_the_e2e_rollback_requires_failed_tests_on_a_main_push(ref, event, tests, cancelled):
+    # "" is what the output reads when e2e-prod never reached the test step,
+    # or was skipped outright. Cancelling the run stops it: an operator cancels
+    # to keep a known flake from reverting a healthy production.
+    allowed = _allowed(
+        "ci.yml",
+        "rollback-production-on-e2e-failure",
+        {"github.ref": ref, "github.event_name": event, "needs.e2e-prod.outputs.tests": tests},
+        cancelled=cancelled,
+    )
+    assert allowed is (ref == "refs/heads/main" and event == "push" and tests == "failure" and not cancelled)