diff --git a/src/v/cluster_link/replication/partition_replicator.cc b/src/v/cluster_link/replication/partition_replicator.cc
index c759f6180c6da..961f2d461922b 100644
--- a/src/v/cluster_link/replication/partition_replicator.cc
+++ b/src/v/cluster_link/replication/partition_replicator.cc
@@ -319,6 +319,7 @@ ss::future<> partition_replicator::fetch_and_replicate() {
 }
 
 ss::future<> partition_replicator::maybe_synchronize_start_offset() {
+    auto shadow_partition_hwm = _sink->high_watermark();
     auto shadow_partition_start_offset = _sink->start_offset();
     auto source_offsets = _source->get_offsets();
 
@@ -333,6 +334,7 @@ ss::future<> partition_replicator::maybe_synchronize_start_offset() {
     }
 
     auto source_start_offset = source_offsets->source_start_offset;
+    auto source_lso = source_offsets->source_lso;
 
     if (source_start_offset <= shadow_partition_start_offset) {
         vlog(
@@ -344,6 +346,29 @@ ss::future<> partition_replicator::maybe_synchronize_start_offset() {
         co_return;
     }
 
+    // The source partition may perform a prefix truncation that lands in the
+    // middle of a batch. Redpanda's data replicators will replicate the whole
+    // batch, including data that starts before the start offset. If we prefix
+    // truncate the shadow partition before that batch is replicated, this will
+    // interfere with our ability to replicate the entire batch, leading to data
+    // loss. To prevent being too eager, we only perform prefix truncation when
+    // we know that we have fully replicated all batches up to or past the
+    // source start offset. This means: source_start_offset <
+    // shadow_partition_hwm OR source_start_offset == source_lso.
+    if (
+      source_lso != source_start_offset
+      && source_start_offset > shadow_partition_hwm) {
+        vlog(
+          _log.debug,
+          "Shadow partition has not replicated up to the source start offset "
+          "yet, deferring prefix truncation. source_start_offset: {}, "
+          "source_lso: {}, shadow_partition_hwm: {}",
+          source_start_offset,
+          source_lso,
+          shadow_partition_hwm);
+        co_return;
+    }
+
     auto truncate_offset = std::max(_start_offset, source_start_offset);
     vlog(
       _log.debug,