diff --git a/runtime/python/prompty/prompty/parsers/prompty.py b/runtime/python/prompty/prompty/parsers/prompty.py
index afc97f111..8a07c7f32 100644
--- a/runtime/python/prompty/prompty/parsers/prompty.py
+++ b/runtime/python/prompty/prompty/parsers/prompty.py
@@ -28,9 +28,15 @@
 
 # Role boundary regex — matches lines like ``system:`` or ``user[name="Alice"]:``
 _ROLE_NAMES = "|".join(sorted(ROLES))
+# Possessive quantifiers + a quoted/unquoted alternation (no shared chars between
+# branches) keep this linear-time — a plain `\w+\s*=\s*"?[^"]*"?` value class lets
+# unquoted attributes overlap with the separator/`]`, causing catastrophic
+# backtracking on malformed input. See issue #446.
+# Possessive quantifiers (`++`, `*+`, `?+`) require Python 3.11+, which pyproject.toml
+# already sets as this package's floor (`requires-python = ">=3.11"`).
 _BOUNDARY_RE = re.compile(
     r"(?im)^\s*#?\s*(" + _ROLE_NAMES + r")"
-    r"(\[((\w+\s*=\s*\"?[^\"]*\"?\s*,?\s*)+)\])?\s*:\s*$"
+    r"(\[(?:\w++\s*+=\s*+(?:\"[^\"]*\"|[^\",\]]*+)\s*+,?+\s*+)+\])?\s*:\s*$"
 )
 
 
diff --git a/runtime/python/prompty/tests/test_parsers.py b/runtime/python/prompty/tests/test_parsers.py
index dfda9c8b0..fb75280d1 100644
--- a/runtime/python/prompty/tests/test_parsers.py
+++ b/runtime/python/prompty/tests/test_parsers.py
@@ -2,6 +2,8 @@
 
 from __future__ import annotations
 
+import time
+
 import pytest
 
 from prompty.core.types import Message, TextPart
@@ -257,6 +259,47 @@ def test_no_images_all_text(self):
         assert isinstance(result[0].parts[0], TextPart)
 
 
+# ---------------------------------------------------------------------------
+# ReDoS regression (issue #446)
+# ---------------------------------------------------------------------------
+
+
+class TestRoleBoundaryReDoS:
+    """Parsing must stay roughly linear in input size on adversarial input.
+
+    Unquoted, unterminated attributes used to let the value class overlap
+    with the `,` separator and closing `]`, giving the backtracking engine
+    an exponential number of equivalent splits to try before failing.
+    The assertion is mostly relative (runtime at 40 reps vs. 20 reps) so it
+    stays robust on a slow/contended CI runner; the small absolute floor
+    only keeps the relative bound from collapsing to noise when `small`
+    itself measures near zero on a very fast run — it is not the budget
+    being enforced.
+    """
+
+    def setup_method(self):
+        self.parser = PromptyChatParser()
+        self.agent = _make_agent()
+
+    def _time_adversarial_parse(self, reps: int, samples: int = 20) -> float:
+        """Best-of-`samples` timing to smooth out scheduler/CPU jitter."""
+        payload = "user[" + "a=b," * reps + "!"
+        best = float("inf")
+        for _ in range(samples):
+            start = time.perf_counter()
+            self.parser.parse(self.agent, payload)
+            best = min(best, time.perf_counter() - start)
+        return best
+
+    def test_doubling_input_does_not_blow_up_runtime(self):
+        small = self._time_adversarial_parse(20)
+        large = self._time_adversarial_parse(40)
+        # Exponential (catastrophic-backtracking) behavior would roughly
+        # square the runtime when the repeat count doubles; linear-time
+        # matching keeps it within a small constant factor.
+        assert large < max(small * 20, 0.1)
+
+
 # ---------------------------------------------------------------------------
 # Async
 # ---------------------------------------------------------------------------