diff --git a/litellm/proxy/pass_through_endpoints/jsonpath_extractor.py b/litellm/proxy/pass_through_endpoints/jsonpath_extractor.py
index 64a8c81624c..8ed54443b7e 100644
--- a/litellm/proxy/pass_through_endpoints/jsonpath_extractor.py
+++ b/litellm/proxy/pass_through_endpoints/jsonpath_extractor.py
@@ -55,7 +55,26 @@ class JsonPathExtractor:
         if not expr or not data:
             return None
 
-        parts: Final = expr.replace("[*]", ".[*]").split(".")
+        # Normalize "key[*]" into "key" + "[*]" segments. Idempotent: the
+        # recursion below passes already-normalized paths (e.g.
+        # "content.[*].text") back through here, and re-running
+        # expr.replace("[*]", ".[*]") on those would inject empty segments
+        # ("content..[*].text") and silently drop every nested-wildcard
+        # expression, skipping the guardrail (#46302).
+        segments: list[str] = []
+        for segment in expr.split("."):
+            if segment == "[*]":
+                segments.append("[*]")
+                continue
+            # Peel trailing wildcards one at a time so adjacent wildcards
+            # ("a[*][*]") and already-normalized paths both split cleanly.
+            while segment.endswith("[*]"):
+                segments.append(segment[:-3])
+                segments.append("[*]")
+                segment = ""
+            if segment:
+                segments.append(segment)
+        parts: Final = segments
         current: Any = data
 
         for i, part in enumerate(parts):
diff --git a/tests/unit/proxy/pass_through_endpoints/test_passthrough_guardrails.py b/tests/unit/proxy/pass_through_endpoints/test_passthrough_guardrails.py
index 40403de846f..56525e3cf72 100644
--- a/tests/unit/proxy/pass_through_endpoints/test_passthrough_guardrails.py
+++ b/tests/unit/proxy/pass_through_endpoints/test_passthrough_guardrails.py
@@ -155,6 +155,26 @@ class TestJsonPathExtractorEvaluate:
         result = JsonPathExtractor.evaluate(data, "messages[*].content")
         assert result == ["You are helpful", "Hello"]
 
+    def test_nested_array_wildcards(self):
+        """Should extract values through two array wildcards.
+
+        Regression test for #46302: the recursive wildcard branch passes the
+        already-normalized remaining path (e.g. "content.[*].text") back
+        through evaluate, where the old replace-based normalization injected
+        an empty segment ("content..[*].text") and silently dropped every
+        nested-wildcard expression - so a pass-through guardrail whose only
+        request_fields entry is messages[*].content[*].text never saw any
+        text and skipped the check entirely.
+        """
+        data = {
+            "messages": [
+                {"role": "user", "content": [{"type": "text", "text": "item1"}]},
+                {"role": "assistant", "content": [{"type": "text", "text": "item2"}]},
+            ]
+        }
+        assert JsonPathExtractor.evaluate(data, "messages[*].content[*].text") == ["item1", "item2"]
+        assert JsonPathExtractor.extract_fields(data, ["messages[*].content[*].text"]) == "item1" + chr(10) + "item2"
+
     def test_missing_key_returns_none(self):
         """Should return None for missing key."""
         data = {"query": "test"}