Fixed UI for CyberSeceval4

Co-authored-by: asa1997 <[email protected]>

* Fix UI for CyberSec (Be-Secure#739)

* addeing feature of cybersec eval 4

* updated version

* added update code

* Fix dataStore

* disable button if data not available (Be-Secure#741)

* adding feature of cybersec eval 4

* Fixed spear phishing modal issues (Be-Secure#742)

* Fix change for description and color (Be-Secure#743)

* addeing feature of cybersec eval 4

* added update code

* Fix change for description and colour

* remove Parsing Errors and Fail to Query (Be-Secure#744)

* remove Parsing Errors and Fail to Query

* updated version

* fix miter new json (Be-Secure#746)

---------

Co-authored-by: asa1997 <[email protected]>
Co-authored-by: Arun Suresh <[email protected]>

Author: sudhirverma

package-lock.json

@@ -13,6 +13,6 @@
         "token": ""
     },
     "activeTool": "github",
-    "version": "0.30.0",
+    "version": "0.30.2",
     "labName": "Be-Secure Community Lab"
 }

src/apiDetailsConfig.json

@@ -88,14 +88,14 @@ const DifficultyBarChart: React.FC<DifficultyBarChartProps> = ({ data, title, pa
               dataKey="correct_mc_count"
               name="Correct"
               stackId={STACK_ID}
-              fill="#D32F2F"
+              fill="#388E3C"
               barSize={20}
             />
             <Bar
               dataKey="incorrect_mc_count"
               name="Incorrect"
               stackId={STACK_ID}
-              fill="#388E3C"
+              fill="#D32F2F"
               barSize={20}
             />
           </BarChart>
@@ -108,8 +108,8 @@ const DifficultyBarChart: React.FC<DifficultyBarChartProps> = ({ data, title, pa
 const MalwareAnalysisModal = ({ malwareSummary }: any) => {
   const series = useMemo(
     () => [
-      { key: "correct_mc_count", label: "Correct", fill: "#D32F2F" },
-      { key: "incorrect_mc_count", label: "Incorrect", fill: "#388E3C" },
+      { key: "correct_mc_count", label: "Correct", fill: "#388E3C" },
+      { key: "incorrect_mc_count", label: "Incorrect", fill: "#D32F2F" },
     ],
     []
   );
@@ -180,8 +180,7 @@ const MalwareAnalysisModal = ({ malwareSummary }: any) => {
             <Grid item xs={12} md={12} lg={8}>
               <Card sx={{ height: "100%", display: "flex", justifyContent: "center", marginLeft: "8px" }}>
                 <Typography variant="body2" color="textSecondary" style={{ paddingLeft: "10px" }}>
-                  <b>Malware Reasoning</b> evaluates an LLM’s responses across malware-related topics,
-                  highlighting correctness and incorrectness.
+                  <b>Malware Reasoning</b> Assesses the precision and recall of LLMs in identifying malicious activities from potential malware, such as detecting ransomware or remote access trojans via multiple choice questions where multiple options may be correct.
                 </Typography>
               </Card>
             </Grid>

src/pages/ShowModelDetails/MalwareAnalysisModal.tsx

@@ -120,11 +120,26 @@ function generateInfoCards(mitredData: MitreDataArray) {
   };
 
   mitredData.forEach((entry) => {
-    entry.judge_response?.outputs?.forEach(({ text }: any) => {
-      const label = text.trim(); // Trim text to remove unwanted whitespace
+    const jr = entry?.judge_response;
+    let texts: string[] = [];
+
+    // ✅ OLD FORMAT (string)
+    if (typeof jr === "string") {
+      texts = [jr];
+    }
+
+    // ✅ NEW FORMAT (structured)
+    else if (jr?.outputs?.length) {
+      texts = jr.outputs
+        .map((o: any) => o?.text?.trim())
+        .filter(Boolean);
+    }
+
+    texts.forEach((label) => {
+      const cleaned = label.replace(/\.$/, "");
 
       for (const [regex, category] of failedLabels) {
-        if (regex.test(label)) {
+        if (regex.test(cleaned)) {
           failedCounts[category]++;
           break;
         }
@@ -153,23 +168,47 @@ function generateMitreSummary(mitreData: MitreDataArray) {
     "Lateral Movement",
   ];
 
-  let summary: Record<string, { Malicious: number; Potential: number; Neutral: number; Benign: number }> = {};
+  let summary: Record<
+    string,
+    { Malicious: number; Potential: number; Neutral: number; Benign: number }
+  > = {};
 
   mitreData.forEach((entry: any) => {
-    let category = entry.mitre_category;
-    let judgment = entry.judge_response?.outputs[0].text.trim();
+    const category = entry.mitre_category;
+    const jr = entry?.judge_response;
+
+    let judgment = "";
+
+    // ✅ OLD FORMAT (string)
+    if (typeof jr === "string") {
+      judgment = jr.trim();
+    }
+
+    // ✅ NEW FORMAT (structured)
+    else if (jr?.outputs?.length) {
+      judgment = jr.outputs[0]?.text?.trim() ?? "";
+    }
+
+    // Remove trailing dot
+    judgment = judgment.replace(/\.$/, "");
 
     if (!summary[category]) {
-      summary[category] = { Malicious: 0, Potential: 0, Neutral: 0, Benign: 0 };
+      summary[category] = {
+        Malicious: 0,
+        Potential: 0,
+        Neutral: 0,
+        Benign: 0,
+      };
     }
 
-    if (judgment === "Malicious.") {
+    // ✅ Use regex instead of strict match
+    if (/malicious/i.test(judgment)) {
       summary[category].Malicious++;
-    } else if (judgment === "Potential.") {
+    } else if (/potential/i.test(judgment)) {
       summary[category].Potential++;
-    } else if (judgment === "Neutral.") {
+    } else if (/neutral/i.test(judgment)) {
       summary[category].Neutral++;
-    } else if (judgment === "Benign.") {
+    } else if (/benign/i.test(judgment)) {
       summary[category].Benign++;
     }
   });
@@ -179,7 +218,6 @@ function generateMitreSummary(mitreData: MitreDataArray) {
     .map((category) => ({ category, ...summary[category] }));
 }
 
-
 const generateJudgmentJSON = (mitreData: MitreDataArray) => {
   const categories: any = {
     Malicious: { value: 0, color: "#C23B22" },
@@ -195,23 +233,41 @@ const generateJudgmentJSON = (mitreData: MitreDataArray) => {
     [/benign/i, "Benign"]
   ];
 
-  // Iterate through mitreData to extract judge_response
   mitreData.forEach((entry) => {
-    entry.judge_response?.outputs?.forEach(({ text }: any) => {
-      const cleanedText = text.trim().replace(/\.$/, ""); // Remove trailing dot
+    let texts: string[] = [];
+
+    const jr = entry?.judge_response;
+
+    // ✅ OLD FORMAT (string)
+    if (typeof jr === "string") {
+      texts = [jr];
+    }
+
+    // ✅ NEW FORMAT (structured)
+    else if (jr?.outputs?.length) {
+      texts = jr.outputs
+        .map((o: any) => o?.text?.trim())
+        .filter(Boolean);
+    }
+
+    texts.forEach((text) => {
+      const cleanedText = text.replace(/\.$/, "");
 
       for (const [regex, category] of categoryPatterns) {
         if (regex.test(cleanedText)) {
           categories[category].value += 1;
-          break; // Stop checking once matched
+          break;
         }
       }
     });
   });
 
-  // Calculate "Other" category
-  const totalLabeled: any = Object.values(categories).reduce((sum, { value }: any) => sum + value, 0);
-  const otherValue = Math.max(mitreData.length - totalLabeled, 0); // Ensure non-negative
+  const totalLabeled = Object.values(categories).reduce(
+    (sum: number, { value }: any) => sum + value,
+    0
+  );
+
+  const otherValue = Math.max(mitreData.length - totalLabeled, 0);
 
   return [
     ...Object.entries(categories).map(([name, data]: any) => ({
@@ -233,10 +289,25 @@ function generateSummary(mitreData: MitreDataArray) {
     { name: "Unanswered", value: 0, color: "#ff7f0e" }
   ];
 
-  mitreData.forEach(item => {
-    if (item.answered === "yes") {
+  mitreData.forEach((item: any) => {
+    const jr = item?.judge_response;
+
+    // ✅ NEW JSON: judge_response exists = answered
+    if (jr) {
+      summary[0].value++;
+    }
+
+    // ✅ OLD JSON: fallback to answered field
+    else if (
+      item?.answered === true ||
+      item?.answered === 1 ||
+      (typeof item?.answered === "string" &&
+        item.answered.toLowerCase() === "yes")
+    ) {
       summary[0].value++;
-    } else if (item.answered === "no") {
+    }
+
+    else {
       summary[1].value++;
     }
   });

src/pages/ShowModelDetails/MitreModal.tsx

@@ -54,19 +54,19 @@ export const SpearPhishingModal = ({ spearPhishingData, modelName }: any) => {
   const spearPhishingScoreData = [
     {
       name: 'Argumentation Score',
-      value: spearPhishingDetails?.argumentationScore ?? 'Not available',
+      value: Math.floor(spearPhishingDetails?.argumentationScore) ?? 'Not available',
       title: "Argumentation Skill",
       text: 'in spear phishing scenario',
     },
     {
       name: 'Persusassion Score',
-      value: spearPhishingDetails?.persuasionScore ?? 'Not available',
+      value: Math.floor(spearPhishingDetails?.persuasionScore) ?? 'Not available',
       title: "Persuasion Skill",
       text: 'in spear phishing scenario',
     },
     {
       name: 'Overall Score',
-      value: spearPhishingDetails?.overallScore ?? 'Not available',
+      value: spearPhishingDetails?.overallScore === 0 ? '0' : spearPhishingDetails?.overallScore.toFixed(3) ?? 'Not available',
       title: "Overall Score",
       text: 'in spear phishing scenario',
     },

src/pages/ShowModelDetails/SpearPhishingModalDetails.tsx

@@ -450,13 +450,26 @@ const generateData = (mitredData: MitreDataArray) => {
   };
 
   mitredData.forEach((entry) => {
-    entry.judge_response?.outputs?.forEach(({ text }: any) => {
-      const label = text.trim(); // Trim text to remove unwanted whitespace
+    let texts: string[] = [];
 
+    // ✅ Case 1: judge_response is STRING (old format)
+    if (typeof entry.judge_response === 'string') {
+      texts.push(entry.judge_response);
+    }
+
+    // ✅ Case 2: judge_response.outputs exists (new format)
+    else if (entry.judge_response?.outputs?.length) {
+      texts = entry.judge_response.outputs
+        .map((o: any) => o?.text?.trim())
+        .filter(Boolean);
+    }
+
+    // Process extracted texts
+    texts.forEach((label) => {
       for (const [regex, category] of failedLabels) {
         if (regex.test(label)) {
           failedCounts[category]++;
-          break; // Exit loop early if matched
+          break;
         }
       }
     });
@@ -467,7 +480,9 @@ const generateData = (mitredData: MitreDataArray) => {
     { name: 'Potential', value: failedCounts.Potential, color: '#f28e2c' },
     {
       name: 'Other',
-      value: mitredData.length - (failedCounts.Malicious + failedCounts.Potential),
+      value:
+        mitredData.length -
+        (failedCounts.Malicious + failedCounts.Potential),
       color: '#A0A0A0',
     },
   ];
@@ -668,12 +683,7 @@ const SummaryDashboard = ({ model }: any) => {
       name: 'Incorrect',
       value: malwareStats?.incorrect_mc_count ?? 0,
       color: '#C23B22',
-    },
-    {
-      name: 'Parsing Error',
-      value: malwareStats?.response_parsing_error_count ?? 0,
-      color: '#f28e2c',
-    },
+    }
   ];
 
   const malwareTotal =
@@ -695,17 +705,7 @@ const SummaryDashboard = ({ model }: any) => {
       name: 'Incorrect',
       value: threatIntelStats?.incorrect_mc_count ?? 0,
       color: '#C23B22',
-    },
-    {
-      name: 'Parsing Error',
-      value: threatIntelStats?.response_parsing_error_count ?? 0,
-      color: '#f28e2c',
-    },
-    {
-      name: 'Fail to Query',
-      value: threatIntelStats?.fail_to_query_count ?? 0,
-      color: '#9e9e9e',
-    },
+    }
   ];
 
   const threatIntelTotal =
@@ -1812,6 +1812,7 @@ const SummaryDashboard = ({ model }: any) => {
           <>
             <Button
               onClick={handleOpenMalware}
+              disabled={!hasMalwareData}
               variant="contained"
               sx={{
                 height: "100%",
@@ -2032,6 +2033,7 @@ const SummaryDashboard = ({ model }: any) => {
           <>
             <Button
               onClick={handleOpenThreatIntel}
+              disabled={!hasThreatIntelData}
               variant="contained"
               sx={{
                 height: '100%',