diff --git a/include/pybind11/pybind11.h b/include/pybind11/pybind11.h
index f57514ae28..83fe1a3d25 100644
--- a/include/pybind11/pybind11.h
+++ b/include/pybind11/pybind11.h
@@ -3356,6 +3356,15 @@ PYBIND11_NOINLINE void keep_alive_impl(handle nurse, handle patient) {
 
 PYBIND11_NOINLINE void
 keep_alive_impl(size_t Nurse, size_t Patient, function_call &call, handle ret) {
+    // The overload bailed out of `load_args` before running, so `ret` is the
+    // PYBIND11_TRY_NEXT_OVERLOAD sentinel ((PyObject *) 1) rather than an object. There is no
+    // call and therefore no relationship to establish; the dispatcher will try the next
+    // overload. Checked here because the sentinel is neither null nor `Py_None`, so it passes
+    // straight through the guards below and is dereferenced.
+    if (ret.ptr() == PYBIND11_TRY_NEXT_OVERLOAD) {
+        return;
+    }
+
     auto get_arg = [&](size_t n) {
         if (n == 0) {
             return ret;
diff --git a/tests/test_call_policies.cpp b/tests/test_call_policies.cpp
index 9140f7e9f2..44a1c0196e 100644
--- a/tests/test_call_policies.cpp
+++ b/tests/test_call_policies.cpp
@@ -9,6 +9,8 @@
 
 #include "pybind11_tests.h"
 
+#include <string>
+
 struct CustomGuard {
     static bool enabled;
 
@@ -110,4 +112,19 @@ TEST_SUBMODULE(call_policies, m) {
     m.def("with_gil", report_gil_status);
     m.def("without_gil", report_gil_status, py::call_guard<py::gil_scoped_release>());
 #endif
+
+    // A keep_alive whose nurse or patient is the return value runs in postcall, which
+    // the dispatcher invokes even when the overload bailed out of load_args. `ret` is then the
+    // PYBIND11_TRY_NEXT_OVERLOAD sentinel rather than an object, and dereferencing it crashed.
+    // Reaching the SECOND overload is what exercises the failed first attempt.
+    struct KeepAliveOverload {};
+    py::class_<KeepAliveOverload>(m, "KeepAliveOverload").def(py::init<>());
+    m.def(
+        "keep_alive_overload",
+        [](KeepAliveOverload *self, int) { return self; },
+        py::keep_alive<0, 1>());
+    m.def(
+        "keep_alive_overload",
+        [](KeepAliveOverload *self, const std::string &) { return self; },
+        py::keep_alive<0, 1>());
 }
diff --git a/tests/test_call_policies.py b/tests/test_call_policies.py
index 11aab9fd9c..4d3937c1c1 100644
--- a/tests/test_call_policies.py
+++ b/tests/test_call_policies.py
@@ -254,3 +254,16 @@ def test_call_guard():
     if hasattr(m, "with_gil"):
         assert m.with_gil() == "GIL held"
         assert m.without_gil() == "GIL released"
+
+
+def test_keep_alive_failed_overload():
+    """A keep_alive on an overload that fails argument conversion must not fire.
+
+    The dispatcher invokes postcall unconditionally, so a keep_alive<0, N> on the
+    overload that bails out of load_args was handed the PYBIND11_TRY_NEXT_OVERLOAD
+    sentinel as its return value and dereferenced it. Calling the second overload is
+    what makes the first one fail first.
+    """
+    obj = m.KeepAliveOverload()
+    assert m.keep_alive_overload(obj, 1) is obj
+    assert m.keep_alive_overload(obj, "x") is obj