diff --git a/lib/lua/vm/stdlib/string.ex b/lib/lua/vm/stdlib/string.ex
index 003cca2..93bb808 100644
--- a/lib/lua/vm/stdlib/string.ex
+++ b/lib/lua/vm/stdlib/string.ex
@@ -217,12 +217,23 @@ defmodule Lua.VM.Stdlib.String do
         # oversized request fails with a catchable error instead of trying to
         # allocate (and OOM the host on) a multi-petabyte binary.
         Limits.check_string_size!(n * byte_size(str) + (n - 1) * byte_size(sep), state.max_string_bytes)
-        Enum.map_join(1..n, sep, fn _ -> str end)
+        build_rep(str, n, sep)
       end
 
     {[result], state}
   end
 
+  # Build the repeated string with allocation proportional to the *result*
+  # size. The obvious `Enum.map_join(1..n, sep, ...)` materializes an
+  # n-element list (plus its iolist) as transient garbage — for a large `n`
+  # that is tens of times the size of the result itself, which spikes the
+  # process heap and trips `max_heap_size` on sandboxed processes (and bloats
+  # the reachable garbage observed through `Lua.call_function!/3`, since that
+  # path returns before the next GC sweeps it). `:binary.copy/2` allocates the
+  # backing binary exactly once, in a single pass.
+  defp build_rep(str, n, ""), do: :binary.copy(str, n)
+  defp build_rep(str, n, sep), do: :binary.copy(str <> sep, n - 1) <> str
+
   # string.reverse(s) - reverses a string byte-by-byte (Lua strings are
   # byte arrays, not codepoint sequences — so a NUL or non-UTF-8 byte
   # must come back at the mirror position).
diff --git a/test/lua/vm/string_test.exs b/test/lua/vm/string_test.exs
index 04ac763..d2581bf 100644
--- a/test/lua/vm/string_test.exs
+++ b/test/lua/vm/string_test.exs
@@ -141,6 +141,47 @@ defmodule Lua.VM.StringTest do
       state = Stdlib.install(State.new())
       assert {:ok, [""], _state} = VM.execute(proto, state)
     end
+
+    # Building the repeated string must allocate proportionally to the result,
+    # not to the repeat count. The previous `Enum.map_join(1..n, ...)`
+    # implementation materialized an n-element list (tens of times the result
+    # size) as transient garbage, spiking the heap and tripping `max_heap_size`
+    # on sandboxed processes for an otherwise modest output. A 16 MB result
+    # comfortably fits a 128 MB cap; the old code blew past 1 GB.
+    test "string.rep allocates proportionally to the result, not the count" do
+      size = 16 * 1024 * 1024
+
+      assert {:ok, result} =
+               heap_capped(128 * 1024 * 1024, fn ->
+                 {[string], _lua} = Lua.eval!(Lua.new(), ~s[return string.rep("x", #{size})])
+                 byte_size(string)
+               end)
+
+      assert result == size
+    end
+  end
+
+  # Runs `fun` in a process whose heap is capped at `byte_limit` (killed if it
+  # exceeds), returning `{:ok, result}` or `:killed`. Lets a test assert that
+  # work stays within a heap budget rather than sampling noisy global memory.
+  defp heap_capped(byte_limit, fun) do
+    parent = self()
+    word_limit = div(byte_limit, :erlang.system_info(:wordsize))
+    cap = %{size: word_limit, kill: true, error_logger: false}
+
+    {pid, ref} =
+      :erlang.spawn_opt(
+        fn -> send(parent, {:result, fun.()}) end,
+        [:monitor, max_heap_size: cap]
+      )
+
+    receive do
+      {:result, result} -> {:ok, result}
+      {:DOWN, ^ref, :process, ^pid, :killed} -> :killed
+      {:DOWN, ^ref, :process, ^pid, reason} -> {:error, reason}
+    after
+      30_000 -> :timeout
+    end
   end
 
   describe "string.reverse" do