BERT-based command injection detection in tool calls (#6599)

2026-01-27 10:46:02 +07:00
parent 64ceea92c8
commit 3dae12765f
5 changed files with 373 additions and 182 deletions
@@ -25,18 +25,13 @@ impl SecurityInspector {
        tool_request_id: String,
    ) -> InspectionResult {
        let action = if security_result.is_malicious && security_result.should_ask_user {
-            // High confidence threat - require user approval with warning
            InspectionAction::RequireApproval(Some(format!(
-                "🔒 Security Alert: This tool call has been flagged as potentially dangerous.\n\
-                Confidence: {:.1}%\n\
-                Explanation: {}\n\
+                "🔒 Security Alert\n\n\
+                {}\n\n\
                Finding ID: {}",
-                security_result.confidence * 100.0,
-                security_result.explanation,
-                security_result.finding_id
+                security_result.explanation, security_result.finding_id
            )))
        } else {
-            // Either not malicious, or below threshold (already logged) - allow
            InspectionAction::Allow
        };