diff --git a/server/app.py b/server/app.py
index 7ad9d59..efd10e9 100644
--- a/server/app.py
+++ b/server/app.py
@@ -146,10 +146,12 @@ def _session_user(request) -> str | None:
 async def csrf_header(request, call_next):
     # ponytail: a cross-site form POST (top-level navigation) does carry the
     # SameSite=Lax cookie, so <form action=.../account/delete> from a
-    # malicious page would work; forms can't set a custom header, the SPA
-    # always sends it. Runs *after* require_auth (registered earlier), so
-    # unauthenticated requests still get 302/401 first.
-    if request.method in ("POST", "DELETE") and request.url.path != "/login":
+    # malicious page would work; forms can't set a custom header, the SPA and
+    # the login page both send one. /login is covered too: login needs no
+    # cookie, so an exempt /login let a cross-site forced-login POST swap a
+    # visitor's session onto the attacker's account. Runs *after* require_auth
+    # (registered earlier), so unauthenticated GETs still 302/401 first.
+    if request.method in ("POST", "DELETE"):
         if request.headers.get("x-aniweb") != "1":
             return JSONResponse({"detail": "forbidden"}, status_code=403)
     return await call_next(request)
diff --git a/tests/test_api.py b/tests/test_api.py
index 407e198..deada4c 100644
--- a/tests/test_api.py
+++ b/tests/test_api.py
@@ -297,7 +297,8 @@ def test_self_delete(base=BASE):
 
 def test_csrf_header_required(base=BASE):
     # mutation routes without the X-Aniweb header -> 403 (a cross-site form
-    # POST can't set it; the SPA always sends it). /login is exempt.
+    # POST can't set it; the SPA always sends it). /login is covered too now
+    # (see test_login_requires_custom_header — forced-login was the hole).
     jar = login("admin", "adminpw12345", base)
     r = urllib.request.Request(base + "/logout", method="POST")
     try:
@@ -320,9 +321,26 @@ def test_stream_token_host_allowlist(base=BASE):
     assert req("GET", f"/stream/{bad2}", jar=jar, base=base) == 400
 
 
+def test_login_requires_custom_header(base=BASE):
+    # forced-login CSRF: /login was the one POST route exempt from the custom
+    # header check, and login needs no cookie — so any cross-site <form>
+    # POST could silently sign a visiting browser into the attacker's account.
+    # /login now demands the header like every other mutation; a cross-site
+    # form cannot set it, so the whole attack class dies at the 403.
+    r = urllib.request.Request(base + "/login", method="POST",
+                               data=json.dumps({"username": "admin",
+                                                "password": "adminpw12345"}).encode(),
+                               headers={"Content-Type": "application/json"})
+    try:
+        urllib.request.build_opener().open(r)
+        raise AssertionError("POST /login without X-Aniweb must be 403")
+    except urllib.error.HTTPError as e:
+        assert e.code == 403, e.code
+
+
 def test_lockout_per_user(base=BASE):
-    # a throwaway name: the lockout is per-username and in-memory, and bob is
-    # also used by the browser tests (a locked-out bob breaks their logins)
+    # a throwaway name: the buckets are in-memory and per-username, and bob
+    # is also used by the browser tests (a locked-out bob breaks their logins)
     codes = [req("POST", "/login", {"username": "lockout-ghost", "password": "bad"}, base=base)
              for _ in range(6)]
     assert codes[:5] == [401] * 5 and codes[5] == 429, codes