diff --git a/src/mcp/server/lowlevel/server.py b/src/mcp/server/lowlevel/server.py
index 25a8fde37..07a51764a 100644
--- a/src/mcp/server/lowlevel/server.py
+++ b/src/mcp/server/lowlevel/server.py
@@ -719,8 +719,10 @@ async def _handle_message(
                     if raise_exceptions:
                         raise message
 
-            for warning in w:  # pragma: no cover
-                logger.info("Warning: %s: %s", warning.category.__name__, warning.message)
+            recorded_warnings = list(w)
+
+        for warning in recorded_warnings:
+            logger.info("Warning: %s: %s", warning.category.__name__, warning.message)
 
     async def _handle_request(
         self,
diff --git a/tests/issues/test_3122_warning_loop.py b/tests/issues/test_3122_warning_loop.py
new file mode 100644
index 000000000..5c8a9a111
--- /dev/null
+++ b/tests/issues/test_3122_warning_loop.py
@@ -0,0 +1,85 @@
+"""Regression test for #3122.
+
+``Server._handle_message`` used to log recorded warnings while still inside the
+``warnings.catch_warnings(record=True)`` block. Because ``record=True`` forces
+an "always" filter, a warning emitted by a logging handler during that logging
+step was appended to the very list being iterated, so the loop never
+terminated. The loop is synchronous, so task cancellation could not interrupt
+it.
+
+The fix snapshots the recorded warnings and logs them after leaving the
+``catch_warnings`` block, so handler-emitted warnings can no longer extend the
+iteration.
+
+The test records two warnings during handling and attaches a logging handler
+that itself warns on the first record only. After the fix each recorded warning
+is logged exactly once, so the handler is invoked twice (once per recorded
+warning) and no extra records appear. Before the fix, the handler's warning fed
+back into the iterated list and the handler was invoked a third time. Warning
+on only the first record keeps the handler from feeding the loop unboundedly, so
+a regressed build terminates and fails the assertion instead of hanging.
+"""
+
+import logging
+import warnings
+from unittest.mock import AsyncMock, Mock
+
+import pytest
+
+import mcp.types as types
+from mcp.server.lowlevel.server import Server
+from mcp.server.session import ServerSession
+from mcp.shared.session import RequestResponder
+
+
+class _WarningEmittingHandler(logging.Handler):
+    """A logging handler whose emit() raises a warning for the first ``cap``
+    records, mimicking e.g. a timestamp formatter that warns per record."""
+
+    def __init__(self, cap: int = 1) -> None:
+        super().__init__()
+        self.emit_count = 0
+        self._cap = cap
+
+    def emit(self, record: logging.LogRecord) -> None:
+        self.emit_count += 1
+        if self.emit_count <= self._cap:
+            warnings.warn("warning raised while logging", stacklevel=1)
+
+
+@pytest.mark.anyio
+async def test_handle_message_logs_each_recorded_warning_once() -> None:
+    server = Server("test-server")
+
+    session = Mock(spec=ServerSession)
+    session.send_log_message = AsyncMock()
+
+    async def _handle_request_that_warns(*args: object, **kwargs: object) -> None:
+        warnings.warn("first warning raised while handling the request", stacklevel=1)
+        warnings.warn("second warning raised while handling the request", stacklevel=1)
+
+    server._handle_request = _handle_request_that_warns  # type: ignore[assignment]
+
+    responder = Mock(spec=RequestResponder)
+    responder.request = types.ClientRequest(root=types.PingRequest(method="ping"))
+    responder.__enter__ = Mock(return_value=responder)
+    responder.__exit__ = Mock(return_value=None)
+
+    server_logger = logging.getLogger("mcp.server.lowlevel.server")
+    handler = _WarningEmittingHandler()
+    server_logger.addHandler(handler)
+    previous_level = server_logger.level
+    server_logger.setLevel(logging.INFO)
+
+    try:
+        with warnings.catch_warnings():
+            warnings.simplefilter("always")
+            await server._handle_message(responder, session, {}, raise_exceptions=False)
+    finally:
+        server_logger.removeHandler(handler)
+        server_logger.setLevel(previous_level)
+
+    # Two warnings were recorded during handling, so the handler is invoked
+    # exactly twice. A regressed build re-appends the handler's own warning and
+    # invokes it a third time.
+    assert handler.emit_count == 2