diff --git a/gems/aws-sdk-s3/CHANGELOG.md b/gems/aws-sdk-s3/CHANGELOG.md
index 49c558bd128..89c093df61c 100644
--- a/gems/aws-sdk-s3/CHANGELOG.md
+++ b/gems/aws-sdk-s3/CHANGELOG.md
@@ -1,6 +1,8 @@
 Unreleased Changes
 ------------------
 
+* Issue - Ensure the source file is closed on multipart `upload_file` part failure, preventing leaked file descriptors (#3408).
+
 1.228.0 (2026-07-16)
 ------------------
 
diff --git a/gems/aws-sdk-s3/lib/aws-sdk-s3/multipart_file_uploader.rb b/gems/aws-sdk-s3/lib/aws-sdk-s3/multipart_file_uploader.rb
index c65f4cd21ff..3f29be384c7 100644
--- a/gems/aws-sdk-s3/lib/aws-sdk-s3/multipart_file_uploader.rb
+++ b/gems/aws-sdk-s3/lib/aws-sdk-s3/multipart_file_uploader.rb
@@ -154,7 +154,6 @@ def upload_with_executor(pending, completed, options)
             Thread.current[:net_http_override_body_stream_chunk] = @http_chunk_size if @http_chunk_size
             update_progress(progress, p)
             resp = @client.upload_part(p)
-            p[:body].close
             completed_part = { etag: resp.etag, part_number: p[:part_number] }
             apply_part_checksum(resp, completed_part)
             completed.push(completed_part)
@@ -162,6 +161,7 @@ def upload_with_executor(pending, completed, options)
             abort_upload = true
             errors << e
           ensure
+            p[:body].close
             Thread.current[:net_http_override_body_stream_chunk] = nil if @http_chunk_size
             completion_queue << :done
           end
diff --git a/gems/aws-sdk-s3/spec/multipart_file_uploader_spec.rb b/gems/aws-sdk-s3/spec/multipart_file_uploader_spec.rb
index 4901ff59361..23aea0c5800 100644
--- a/gems/aws-sdk-s3/spec/multipart_file_uploader_spec.rb
+++ b/gems/aws-sdk-s3/spec/multipart_file_uploader_spec.rb
@@ -125,6 +125,37 @@ module S3
           expect { subject.upload(large_file, params) }.to raise_error(/multipart upload failed: part 3 failed/)
         end
 
+        it 'closes all file parts even when a part upload fails' do
+          # Fail the last part so the posting loop can't break early and skip
+          # un-posted parts. This keeps the assertion deterministic across MRI
+          # (GIL-serialized) and JRuby (truly parallel) threads.
+          file = Tempfile.new('six-meg-file').tap do |f|
+            6.times { f.write(one_mb) }
+            f.rewind
+          end
+          file_parts = []
+          allow(FilePart).to receive(:new).and_wrap_original do |original, *args|
+            original.call(*args).tap do |fp|
+              allow(fp).to receive(:close).and_call_original
+              file_parts << fp
+            end
+          end
+
+          client.stub_responses(
+            :upload_part,
+            [
+              { etag: 'etag-1' },
+              RuntimeError.new('part 2 failed')
+            ]
+          )
+
+          expect { subject.upload(file, params) }
+            .to raise_error(/multipart upload failed: part 2 failed/)
+
+          expect(file_parts.size).to eq(2)
+          file_parts.each { |fp| expect(fp).to have_received(:close) }
+        end
+
         it 'reports when it is unable to abort a failed multipart upload', :jruby_flaky do
           client.stub_responses(:upload_part, RuntimeError.new('part failed'))
           client.stub_responses(:abort_multipart_upload, RuntimeError.new('network-error'))