diff --git a/kernels/portable/cpu/op_convolution.cpp b/kernels/portable/cpu/op_convolution.cpp
index bf848bc53c2..efb6e5898cf 100644
--- a/kernels/portable/cpu/op_convolution.cpp
+++ b/kernels/portable/cpu/op_convolution.cpp
@@ -203,6 +203,15 @@ void convolution_wrapper(
     bool transposed,
     int64_t groups,
     Tensor& out) {
+  // Defense-in-depth: get_load_to_compute_fn returns nullptr (and sets
+  // ctx.fail()) when bias' scalar_type is outside the dispatcher's
+  // supported set. check_convolution_args() in kernel_ops_util.cpp now
+  // rejects such bias tensors at the trust boundary, but we keep this guard
+  // so any future caller of convolution_wrapper is safe; load_bias is
+  // dereferenced inside the wrapper / conv2d_impl whenever bias has data.
+  if (bias.has_value() && load_bias == nullptr) {
+    return;
+  }
   SizesArrayRef in_sizes = in.sizes();
   SizesArrayRef weight_sizes = weight.sizes();
   SizesArrayRef out_sizes = out.sizes();
diff --git a/kernels/portable/cpu/op_cumsum.cpp b/kernels/portable/cpu/op_cumsum.cpp
index 3a518d30715..2ad35941c8b 100644
--- a/kernels/portable/cpu/op_cumsum.cpp
+++ b/kernels/portable/cpu/op_cumsum.cpp
@@ -44,6 +44,14 @@ void cumsum_tensors(
   if (self.numel() == 0) {
     return;
   }
+  // Defense-in-depth: get_load_to_compute_fn returns nullptr (and sets
+  // ctx.fail()) when the input tensor's scalar_type is outside the dispatcher's
+  // supported set. check_cumsum_args() in kernel_ops_util.cpp now rejects such
+  // inputs at the trust boundary, but we keep this guard so any future caller
+  // of cumsum_tensors is safe.
+  if (load_self == nullptr) {
+    return;
+  }
 
   const char* const input_data_base =
       reinterpret_cast<const char*>(self.const_data_ptr());
diff --git a/kernels/portable/cpu/util/kernel_ops_util.cpp b/kernels/portable/cpu/util/kernel_ops_util.cpp
index 2e0ceee07e5..0bc20dcd5a2 100644
--- a/kernels/portable/cpu/util/kernel_ops_util.cpp
+++ b/kernels/portable/cpu/util/kernel_ops_util.cpp
@@ -380,6 +380,8 @@ bool check_convolution_args(
     int64_t groups,
     const Tensor& out) {
   ET_LOG_AND_RETURN_IF_FALSE(tensors_have_same_dtype(in, weight, out));
+  ET_LOG_AND_RETURN_IF_FALSE(
+      executorch::runtime::tensor_is_realhbf16_type(in));
 
   ET_LOG_AND_RETURN_IF_FALSE(tensor_is_default_or_channels_last_dim_order(in));
   ET_LOG_AND_RETURN_IF_FALSE(
@@ -395,6 +397,8 @@ bool check_convolution_args(
 
   if (bias.has_value()) {
     ET_LOG_AND_RETURN_IF_FALSE(tensor_is_rank(bias.value(), 1));
+    ET_LOG_AND_RETURN_IF_FALSE(
+        executorch::runtime::tensor_is_realhbf16_type(bias.value()));
     ET_CHECK_OR_RETURN_FALSE(
         bias.value().size(0) == transposed ? groups * weight.size(1)
                                            : weight.size(0),
@@ -510,6 +514,10 @@ bool check_cumsum_args(
     optional<ScalarType> dtype,
     Tensor& out) {
   ET_LOG_AND_RETURN_IF_FALSE(dim_is_valid(dim, in.dim()));
+  ET_LOG_AND_RETURN_IF_FALSE(
+      executorch::runtime::tensor_is_realhbbf16_type(in));
+  ET_LOG_AND_RETURN_IF_FALSE(
+      executorch::runtime::tensor_is_realhbbf16_type(out));
 
   if (dtype.has_value()) {
     ET_LOG_AND_RETURN_IF_FALSE(dtype.value() == out.scalar_type());