diff --git a/include/nvexec/detail/variant.cuh b/include/nvexec/detail/variant.cuh
index 16bb64ab5..0ec55b386 100644
--- a/include/nvexec/detail/variant.cuh
+++ b/include/nvexec/detail/variant.cuh
@@ -192,7 +192,7 @@ namespace nv::execution
     }
 
     union_t storage_;
-    index_t index_;
+    index_t index_{detail::npos<index_t>()};
   };
 }  // namespace nv::execution
 
diff --git a/test/nvexec/variant.cpp b/test/nvexec/variant.cpp
index 610d1cf9e..fd63181f8 100644
--- a/test/nvexec/variant.cpp
+++ b/test/nvexec/variant.cpp
@@ -20,6 +20,9 @@
 #include <cuda/std/tuple>
 #include <thrust/universal_vector.h>
 
+#include <cstddef>
+#include <new>
+
 #include "common.cuh"
 #include "nvexec/detail/variant.cuh"
 
@@ -31,6 +34,19 @@ using nvexec::visit;
 
 namespace
 {
+  int variant_destructor_calls = 0;
+
+  struct variant_destructor_probe
+  {
+    STDEXEC_ATTRIBUTE(host, device)
+    variant_destructor_probe() = default;
+
+    STDEXEC_ATTRIBUTE(host, device)
+    ~variant_destructor_probe()
+    {
+      NV_IF_TARGET(NV_IS_HOST, (++variant_destructor_calls;));
+    }
+  };
 
   TEST_CASE("nvexec variant max size is correct", "[cuda][stream][containers][variant]")
   {
@@ -57,6 +73,21 @@ namespace
     STATIC_REQUIRE(variant_t<char, int, double>::size == 3);
   }
 
+  TEST_CASE("nvexec variant initializes its index before constructing the first alternative",
+            "[cuda][stream][containers][variant]")
+  {
+    using variant_type = variant_t<variant_destructor_probe>;
+
+    alignas(variant_type) std::byte storage[sizeof(variant_type)]{};
+    variant_destructor_calls = 0;
+
+    auto* variant = ::new (storage) variant_type;
+    CHECK(variant_destructor_calls == 0);
+
+    variant->~variant_type();
+    CHECK(variant_destructor_calls == 1);
+  }
+
   TEST_CASE("nvexec variant emplaces alternative from CPU", "[cuda][stream][containers][variant]")
   {
     variant_t<int, double> v;