diff --git a/src/plugins/intel_gpu/src/graph/crop.cpp b/src/plugins/intel_gpu/src/graph/crop.cpp
index 146a1fa89b400b..09c5f01f216e57 100644
--- a/src/plugins/intel_gpu/src/graph/crop.cpp
+++ b/src/plugins/intel_gpu/src/graph/crop.cpp
@@ -250,7 +250,7 @@ crop_inst::typed_primitive_inst(network& network, crop_node const& node) : paren
                                         "Invalid Batch offset: exceeds data for output!");
     }
 
-    if (node.can_be_optimized()) {
+    if (!node.is_dynamic() && node.can_be_optimized()) {
         update_output_memory();
     }
 }
diff --git a/src/plugins/intel_gpu/src/graph/graph_optimizer/prepare_buffer_fusing.cpp b/src/plugins/intel_gpu/src/graph/graph_optimizer/prepare_buffer_fusing.cpp
index e3471b37c05bd9..7f1fb69446edb9 100644
--- a/src/plugins/intel_gpu/src/graph/graph_optimizer/prepare_buffer_fusing.cpp
+++ b/src/plugins/intel_gpu/src/graph/graph_optimizer/prepare_buffer_fusing.cpp
@@ -500,6 +500,14 @@ bool crop_in_place_optimization::match(const program_node& node,
     if (node.is_constant())
         return false;
 
+    // do not optimize variadic_split crop when either input1 or input2 is not constant.
+    // VariadicSplit ngraph shape infer requires value of axis(input1) and split_lengths(input2).
+    // And non_constant input1/input2 makes risky execution of runtime buffer fusing.
+    auto& crop_node = node.as<crop>();
+    if ((crop_node.get_primitive()->op_mode == cldnn::crop_ngraph_op_mode::variadic_split) &&
+        (!crop_node.get_dependency(1).is_constant() || !crop_node.get_dependency(2).is_constant()))
+        return false;
+
     if (node.get_users().size() > 0) {
         if (node.get_program().is_body_program() && node.get_dependency(0).is_type<lstm_elt>()) {
             return false;
diff --git a/src/plugins/intel_gpu/tests/unit/passes/prepare_buffer_fusing_test.cpp b/src/plugins/intel_gpu/tests/unit/passes/prepare_buffer_fusing_test.cpp
index e4a077594c7a7e..e5506388eba273 100644
--- a/src/plugins/intel_gpu/tests/unit/passes/prepare_buffer_fusing_test.cpp
+++ b/src/plugins/intel_gpu/tests/unit/passes/prepare_buffer_fusing_test.cpp
@@ -777,6 +777,91 @@ TEST(prepare_buffer_fusing, in_place_crop_dynamic) {
         ASSERT_EQ(output_ptr_3[i], out3[i]);
 }
 
+TEST(prepare_buffer_fusing, in_place_crop_dynamic_split_lengths) {
+    auto& engine = get_test_engine();
+
+    auto in_layout = layout{ ov::PartialShape{-1, -1, -1}, data_types::f32, format::bfyx};
+    auto in2_layout = layout{ ov::PartialShape{-1, -1}, data_types::f32, format::bfyx};
+    auto input_mem = engine.allocate_memory({ {1, 2, 4}, data_types::f32, format::bfyx });
+    auto weights_mem = engine.allocate_memory({ {8, 4}, data_types::u8, format::bfyx });
+    auto bias_mem = engine.allocate_memory({ {1, 1, 8}, data_types::f32, format::bfyx });
+    auto scale_mem = engine.allocate_memory({ {8, 1}, data_types::f32, format::bfyx });
+    auto zp_mem = engine.allocate_memory({ {8, 1}, data_types::f32, format::bfyx });
+    auto axis_mem = engine.allocate_memory({ {}, data_types::i64, format::bfyx });
+    auto shapeof_mem = engine.allocate_memory({ {2, 6}, data_types::f32, format::bfyx });
+
+    int64_t axis = 2;
+    set_values(input_mem, { -0.5f,  2.0f,  0.5f,  1.0f,
+                             0.5f, -2.0f, -0.5f, -1.0f });
+    set_values<int64_t>(axis_mem, {axis});
+    set_values(shapeof_mem, { 1.0f,  2.0f,  3.0f,  4.0f,
+                              5.0f,  6.0f,  7.0f,  8.0f,
+                              9.0f, 10.0f, 11.0f, 12.0f});
+    set_values<uint8_t>(weights_mem, { 1,  2,  3,  4,
+                                       5,  6,  7,  8,
+                                       9, 10, 11, 12,
+                                      13, 14, 15,  0,
+                                      15, 14, 13, 12,
+                                      11, 10,  9,  8,
+                                       7,  6,  5,  4,
+                                       3,  2,  1,  0});
+    set_values(bias_mem, { 1.0f, -2.0f, 3.0f, -4.0f, 5.0f, -6.0f, 7.0f, 2.0f });
+    set_values(scale_mem, { 2.0f, 4.0f, -2.0f, -4.0f, 0.5f, -0.5f, 2.0f, 2.0f });
+    set_values(zp_mem, { 1.0f, 2.0f, 2.0f, 1.0f, 4.0f, 1.0f, 6.0f, 2.0f });
+
+    std::vector<float> out1 = { 13.f, 58.f, -11.f, -62.f };
+    std::vector<float> out2 = { -51.f, -108.f, 18.5f, -18.f, 1.f, -4.f, 57.f, 100.f, -8.5f, 6.f, 13.f, 8.f };
+    std::vector<float> out3 = { 13.f, 58.f, -51.f, -108.f, 18.5f, -18.f, 1.f, -4.f, -11.f, -62.f, 57.f, 100.f, -8.5f, 6.f, 13.f, 8.f };
+
+    cldnn::crop_ngraph_op_mode op_mode = cldnn::crop_ngraph_op_mode::variadic_split;
+    topology topology(
+        input_layout("input", in_layout),
+        input_layout("input_shapeof", in2_layout),
+        data("axis", axis_mem),
+        data("weights", weights_mem),
+        data("bias", bias_mem),
+        data("scale", scale_mem),
+        data("zp", zp_mem),
+        fully_connected("fc", input_info("input"), "weights", "bias", "scale", "zp", data_types::f32, 3, 2),
+        shape_of("shapeof", input_info("input_shapeof"), cldnn::data_types::i64),
+        crop("crop1", { input_info("fc"), input_info("axis"), input_info("shapeof") }, cldnn::tensor(1), cldnn::tensor(0), op_mode, 0, axis),
+        reorder("output1", input_info("crop1"), format::bfyx, data_types::f32),
+        crop("crop2", { input_info("fc"), input_info("axis"), input_info("shapeof") }, cldnn::tensor(1), cldnn::tensor(0), op_mode, 1, axis),
+        reshape("reshape", input_info("crop2"), true, std::vector<int64_t>{0, 0, 3, 2}, ov::PartialShape{-1, -1, 3, 2}, cldnn::reshape::reshape_mode::base),
+        reorder("output2", input_info("reshape"), format::bfyx, data_types::f32, std::vector<float>(), reorder_mean_mode::subtract, padding(), true),
+        reorder("output3", input_info("fc"), format::bfyx, data_types::f32)
+    );
+
+    auto config = get_test_default_config(engine);
+    config.set_property(ov::intel_gpu::allow_new_shape_infer(true));
+    config.set_property(ov::intel_gpu::optimize_data(true));
+    network network(engine, topology, config);
+
+    network.set_input_data("input", input_mem);
+    network.set_input_data("input_shapeof", shapeof_mem);
+
+    std::map<cldnn::primitive_id, cldnn::network_output> outputs;
+    EXPECT_NO_THROW(outputs = network.execute());
+
+    auto output = outputs.at("output1").get_memory();
+    cldnn::mem_lock<float> output_ptr(output, get_test_stream());
+
+    for (size_t i = 0; i < out1.size(); i++)
+        ASSERT_EQ(output_ptr[i], out1[i]);
+
+    auto output_2 = outputs.at("output2").get_memory();
+    cldnn::mem_lock<float> output_ptr_2(output_2, get_test_stream());
+
+    for (size_t i = 0; i < out2.size(); i++)
+        ASSERT_EQ(output_ptr_2[i], out2[i]);
+
+    auto output_3 = outputs.at("output3").get_memory();
+    cldnn::mem_lock<float> output_ptr_3(output_3, get_test_stream());
+
+    for (size_t i = 0; i < out3.size(); i++)
+        ASSERT_EQ(output_ptr_3[i], out3[i]);
+}
+
 // Testing for implicit crop along batch axis and outer padding optimzing.
 // Outer padding opt includes opt out of reshape and reorder which has padded input only in batch axis
 // This optimzing also includes offset(outer axis padded input) handling of oneDNN primitive.