diff --git a/crates/dbx-core/src/db/mysql.rs b/crates/dbx-core/src/db/mysql.rs
index 7e5e5989a3..428959072c 100644
--- a/crates/dbx-core/src/db/mysql.rs
+++ b/crates/dbx-core/src/db/mysql.rs
@@ -1468,7 +1468,11 @@ fn mysql_top_level_limit(sql: &str) -> Option<usize> {
         if depth == 0 && mysql_keyword_at(sql, i, "LIMIT") {
             return parse_mysql_limit_value(sql, i + "LIMIT".len());
         }
+        // Move to next byte, but ensure we stay on a UTF-8 boundary
         i += 1;
+        while i < bytes.len() && !sql.is_char_boundary(i) {
+            i += 1;
+        }
     }
 
     None
@@ -1497,13 +1501,20 @@ fn parse_usize_token(sql: &str, i: &mut usize) -> Option<usize> {
     if *i == start {
         return None;
     }
-    sql[start..*i].parse().ok()
+    // Ensure the slice is valid UTF-8 before parsing
+    std::str::from_utf8(&bytes[start..*i]).ok()?.parse().ok()
 }
 
 fn mysql_keyword_at(sql: &str, i: usize, keyword: &str) -> bool {
     let end = i + keyword.len();
-    end <= sql.len()
-        && sql[i..end].eq_ignore_ascii_case(keyword)
+    if end > sql.len() {
+        return false;
+    }
+    // Ensure indices are on UTF-8 boundaries before slicing
+    if !sql.is_char_boundary(i) || !sql.is_char_boundary(end) {
+        return false;
+    }
+    sql[i..end].eq_ignore_ascii_case(keyword)
         && (i == 0 || !is_mysql_identifier_byte(sql.as_bytes()[i - 1]))
         && (end == sql.len() || !is_mysql_identifier_byte(sql.as_bytes()[end]))
 }