diff --git a/scripts/lib/mlkem_test.py b/scripts/lib/mlkem_test.py
index 60b12fb4f..590dabac9 100644
--- a/scripts/lib/mlkem_test.py
+++ b/scripts/lib/mlkem_test.py
@@ -447,9 +447,8 @@ def bench(
     def all(self, func: bool, kat: bool, nistkat: bool):
         config_logger(self.verbose)
 
-        exit_code = 0
-
         def all(opt: bool):
+            code = 0
             if self.compile:
                 compiles = [
                     *([self._func.compile] if func else []),
@@ -461,7 +460,7 @@ def all(opt: bool):
                     try:
                         f(opt)
                     except SystemExit as e:
-                        exit_code = exit_code or e
+                        code = code or e
 
                     sys.stdout.flush()
 
@@ -474,15 +473,18 @@ def all(opt: bool):
 
                 for f in runs:
                     try:
-                        f(opt)
+                        code = code or int(f(opt))
                     except SystemExit as e:
-                        exit_code = exit_code or e
+                        code = code or e
 
                     sys.stdout.flush()
+            return code
+
+        exit_code = 0
 
         if self.opt.lower() == "all" or self.opt.lower() == "no_opt":
-            all(False)
+            exit_code = exit_code or all(False)
         if self.opt.lower() == "all" or self.opt.lower() == "opt":
-            all(True)
+            exit_code = exit_code or all(True)
 
         exit(exit_code)
diff --git a/scripts/lib/util.py b/scripts/lib/util.py
index 993ec4ff1..d92efe4af 100644
--- a/scripts/lib/util.py
+++ b/scripts/lib/util.py
@@ -159,7 +159,9 @@ def add_summaries(fn, title, summaries):
 
 
 def config_logger(verbose):
-    logging.basicConfig(format="%(levelname)-5s > %(name)-40s %(message)s")
+    logging.basicConfig(
+        stream=sys.stdout, format="%(levelname)-5s > %(name)-40s %(message)s"
+    )
 
     logger = logging.getLogger()