From 8288e351ef203bb4ea28577aaf7b7374771697e5 Mon Sep 17 00:00:00 2001
From: Sameer Kankute <sameer@berri.ai>
Date: Tue, 24 Feb 2026 09:14:10 +0530
Subject: [PATCH] Fix: test_sentence[dt_8]

---
 .../litellm_content_filter/content_filter.py  | 21 ++++++++++++++++++-
 .../sg_mas_human_oversight.yaml               |  8 +++----
 .../sg_mas_transparency_explainability.yaml   |  3 ++-
 3 files changed, 26 insertions(+), 6 deletions(-)

diff --git a/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py b/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py
index d569e69e56..d0ad9eb404 100644
--- a/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py
+++ b/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py
@@ -214,6 +214,10 @@ class ContentFilterGuardrail(CustomGuardrail):
         self.category_keywords: Dict[str, Tuple[str, str, ContentFilterAction]] = (
             {}
         )  # keyword -> (category, severity, action)
+        # Always-block keywords are checked before exceptions (cannot be bypassed)
+        self.always_block_category_keywords: Dict[
+            str, Tuple[str, str, ContentFilterAction]
+        ] = {}
         # Store conditional categories (identifier_words + block_words)
         self.conditional_categories: Dict[str, Dict[str, Any]] = (
             {}
@@ -452,7 +456,7 @@ class ContentFilterGuardrail(CustomGuardrail):
                         keyword = keyword_data["keyword"].lower()
                         severity = keyword_data.get("severity", "high")
                         if self._should_apply_severity(severity, severity_threshold):
-                            self.category_keywords[keyword] = (
+                            self.always_block_category_keywords[keyword] = (
                                 category_name,
                                 severity,
                                 category_action,
@@ -1064,6 +1068,20 @@ class ContentFilterGuardrail(CustomGuardrail):
         """
         text_lower = text.lower()
 
+        # Always-block keywords are matched first — exceptions do not apply to them.
+        for keyword, (category, severity, action) in self.always_block_category_keywords.items():
+            keyword_pattern_str = self._keyword_to_regex_pattern(keyword)
+            if " " in keyword:
+                keyword_found = bool(re.search(keyword_pattern_str, text_lower))
+            else:
+                keyword_pattern = r"\b" + keyword_pattern_str + r"\b"
+                keyword_found = bool(re.search(keyword_pattern, text_lower))
+            if keyword_found:
+                verbose_proxy_logger.debug(
+                    f"Always-block keyword '{keyword}' found in category '{category}'"
+                )
+                return (keyword, category, severity, action)
+
         # First check if any exception applies
         for exception in exceptions:
             if exception in text_lower:
@@ -1571,6 +1589,7 @@ class ContentFilterGuardrail(CustomGuardrail):
             len(self.compiled_patterns)
             + len(self.blocked_words)
             + len(self.category_keywords)
+            + len(self.always_block_category_keywords)
         )
 
     def _get_policy_templates(self) -> Optional[str]:
diff --git a/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/sg_mas_human_oversight.yaml b/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/sg_mas_human_oversight.yaml
index dfca04f901..0e9e336dbe 100644
--- a/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/sg_mas_human_oversight.yaml
+++ b/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/sg_mas_human_oversight.yaml
@@ -74,16 +74,16 @@ always_block_keywords:
 exceptions:
   - "human in the loop"
   - "human on the loop"
-  - "human review"
-  - "human oversight"
+  - "add human review"
+  - "with human review"
   - "manual review"
   - "escalation"
   - "appeals process"
   - "override mechanism"
-  - "explain"
+  - "explain what"
+  - "explain how"
   - "what is"
   - "research"
   - "improve oversight"
-  - "add human review"
   - "compliance"
   - "governance"
diff --git a/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/sg_mas_transparency_explainability.yaml b/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/sg_mas_transparency_explainability.yaml
index 19b999729b..0a08abd86e 100644
--- a/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/sg_mas_transparency_explainability.yaml
+++ b/litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/sg_mas_transparency_explainability.yaml
@@ -73,7 +73,8 @@ exceptions:
   - "interpretability"
   - "model card"
   - "audit trail"
-  - "explain"
+  - "explain what"
+  - "explain how"
   - "what is"
   - "SHAP"
   - "LIME"