Resolve dot-segments when deriving and matching OAuth resource URLs

src/mcp/shared/auth_utils.py

@@ -1,16 +1,41 @@
 """Utilities for OAuth 2.0 Resource Indicators (RFC 8707) and PKCE (RFC 7636)."""
 
 import time
-from urllib.parse import urlparse, urlsplit, urlunsplit
+from urllib.parse import urlsplit, urlunsplit
 
 from pydantic import AnyUrl, HttpUrl
 
+# WHATWG URL treats these percent-encoded spellings as dot-segments too.
+_SINGLE_DOT_SEGMENTS = {".", "%2e"}
+_DOUBLE_DOT_SEGMENTS = {"..", ".%2e", "%2e.", "%2e%2e"}
+
+
+def _remove_dot_segments(path: str) -> str:
+    """Resolve "." and ".." segments in a URL path (RFC 3986 section 5.2.4)."""
+    segments = path.split("/")
+    output: list[str] = []
+    for index, segment in enumerate(segments):
+        is_last = index == len(segments) - 1
+        kind = segment.lower()
+        if kind in _DOUBLE_DOT_SEGMENTS:
+            if len(output) > 1:
+                output.pop()
+            if is_last:
+                output.append("")
+        elif kind in _SINGLE_DOT_SEGMENTS:
+            if is_last:
+                output.append("")
+        else:
+            output.append(segment)
+    return "/".join(output)
+
 
 def resource_url_from_server_url(url: str | HttpUrl | AnyUrl) -> str:
     """Convert server URL to canonical resource URL per RFC 8707.
 
     RFC 8707 section 2 states that resource URIs "MUST NOT include a fragment component".
-    Returns absolute URI with lowercase scheme/host for canonical form.
+    Returns absolute URI with lowercase scheme/host and dot-segments resolved, so the
+    resource identifies the same location an HTTP client would actually request.
 
     Args:
         url: Server URL to convert
@@ -23,7 +48,14 @@ def resource_url_from_server_url(url: str | HttpUrl | AnyUrl) -> str:
 
     # Parse the URL and remove fragment, create canonical form
     parsed = urlsplit(url_str)
-    canonical = urlunsplit(parsed._replace(scheme=parsed.scheme.lower(), netloc=parsed.netloc.lower(), fragment=""))
+    canonical = urlunsplit(
+        parsed._replace(
+            scheme=parsed.scheme.lower(),
+            netloc=parsed.netloc.lower(),
+            path=_remove_dot_segments(parsed.path),
+            fragment="",
+        )
+    )
 
     return canonical
 
@@ -34,7 +66,8 @@ def check_resource_allowed(requested_resource: str, configured_resource: str) ->
     A requested resource matches if it has the same scheme, domain, port,
     and its path starts with the configured resource's path. This allows
     hierarchical matching where a token for a parent resource can be used
-    for child resources.
+    for child resources. Dot-segments in either path are resolved before
+    comparing.
 
     Args:
         requested_resource: The resource URL being requested
@@ -44,17 +77,17 @@ def check_resource_allowed(requested_resource: str, configured_resource: str) ->
         True if the requested resource matches the configured resource
     """
     # Parse both URLs
-    requested = urlparse(requested_resource)
-    configured = urlparse(configured_resource)
+    requested = urlsplit(requested_resource)
+    configured = urlsplit(configured_resource)
 
     # Compare scheme, host, and port (origin)
     if requested.scheme.lower() != configured.scheme.lower() or requested.netloc.lower() != configured.netloc.lower():
         return False
 
     # Normalize trailing slashes before comparison so that
     # "/foo" and "/foo/" are treated as equivalent.
-    requested_path = requested.path
-    configured_path = configured.path
+    requested_path = _remove_dot_segments(requested.path)
+    configured_path = _remove_dot_segments(configured.path)
     if not requested_path.endswith("/"):
         requested_path += "/"
     if not configured_path.endswith("/"):

tests/client/test_auth.py

@@ -862,6 +862,36 @@ async def test_validate_resource_rejects_mismatched_resource(
         await provider._validate_resource_match(prm)
 
 
+@pytest.mark.anyio
+async def test_validate_resource_rejects_sibling_path_reached_via_dot_segments(
+    client_metadata: OAuthClientMetadata, mock_storage: MockTokenStorage
+) -> None:
+    """A `server_url` whose dot-segments resolve to `/m/mcp` rejects a PRM `resource` of `/victim/mcp`.
+
+    SDK-defined: the resource identifier is derived from the location the HTTP client actually
+    requests (RFC 3986 section 5.2.4), so a same-origin sibling path is neither accepted during
+    discovery nor adopted as the RFC 8707 `resource` parameter.
+    """
+    provider = OAuthClientProvider(
+        server_url="https://shared.example.com/victim/mcp/../../m/mcp",
+        client_metadata=client_metadata,
+        storage=mock_storage,
+    )
+    provider._initialized = True
+
+    prm = ProtectedResourceMetadata(
+        resource=AnyHttpUrl("https://shared.example.com/victim/mcp"),
+        authorization_servers=[AnyHttpUrl("https://auth.example.com")],
+    )
+    with pytest.raises(OAuthFlowError) as exc_info:
+        await provider._validate_resource_match(prm)
+    assert str(exc_info.value) == snapshot(
+        "Protected resource https://shared.example.com/victim/mcp does not match expected https://shared.example.com/m/mcp"
+    )
+    provider.context.protected_resource_metadata = prm
+    assert provider.context.get_resource_url() == snapshot("https://shared.example.com/m/mcp")
+
+
 @pytest.mark.anyio
 async def test_validate_resource_accepts_matching_resource(
     client_metadata: OAuthClientMetadata, mock_storage: MockTokenStorage

tests/shared/test_auth_utils.py

@@ -1,6 +1,9 @@
 """Tests for OAuth 2.0 Resource Indicators utilities."""
 
-from pydantic import HttpUrl
+import itertools
+
+import pytest
+from pydantic import AnyHttpUrl, HttpUrl
 
 from mcp.shared.auth_utils import check_resource_allowed, resource_url_from_server_url
 
@@ -46,6 +49,42 @@ def test_resource_url_from_server_url_handles_pydantic_urls():
     assert resource_url_from_server_url(url) == "https://example.com/path"
 
 
+@pytest.mark.parametrize(
+    ("server_url", "expected"),
+    [
+        ("https://example.com/api/../admin", "https://example.com/admin"),
+        ("https://example.com/api/%2E%2e/admin", "https://example.com/admin"),
+        ("https://example.com/api/.%2e/admin", "https://example.com/admin"),
+        ("https://example.com/api/./v1", "https://example.com/api/v1"),
+        ("https://example.com/api/v1/..", "https://example.com/api/"),
+        ("https://example.com/api/v1/.", "https://example.com/api/v1/"),
+        ("https://example.com/../admin", "https://example.com/admin"),
+        ("https://example.com/a//b", "https://example.com/a//b"),
+        ("https://example.com/a%2Fb/c", "https://example.com/a%2Fb/c"),
+    ],
+)
+def test_resource_url_from_server_url_resolves_dot_segments(server_url: str, expected: str):
+    """Dot-segments (including `%2E` spellings) are resolved per RFC 3986 section 5.2.4.
+
+    Empty segments and encoded slashes are not path separators and stay as written.
+    """
+    assert resource_url_from_server_url(server_url) == expected
+
+
+def test_resource_url_from_server_url_path_matches_whatwg_resolution_for_literal_dot_segments():
+    """Every combination of literal `.`, `..`, empty and plain segments resolves as pydantic's WHATWG parser does.
+
+    The PRM `resource` side is parsed by `AnyHttpUrl`, so both operands of `check_resource_allowed`
+    must agree on dot-segment resolution for the comparison to be meaningful.
+    """
+    atoms = ["", ".", "..", "a", "b.", "..."]
+    for count in range(1, 5):
+        for segments in itertools.product(atoms, repeat=count):
+            path = "/" + "/".join(segments)
+            expected = AnyHttpUrl(f"https://example.com{path}").path
+            assert resource_url_from_server_url(f"https://example.com{path}") == f"https://example.com{expected}"
+
+
 # Tests for check_resource_allowed function
 
 
@@ -121,3 +160,30 @@ def test_check_resource_allowed_empty_paths():
     assert check_resource_allowed("https://example.com", "https://example.com") is True
     assert check_resource_allowed("https://example.com/", "https://example.com") is True
     assert check_resource_allowed("https://example.com/api", "https://example.com") is True
+
+
+@pytest.mark.parametrize(
+    "requested",
+    [
+        "https://example.com/api/../admin",
+        "https://example.com/api/%2e%2e/admin",
+        "https://example.com/api/v1/../../admin",
+        "https://example.com/api/..",
+    ],
+)
+def test_check_resource_allowed_rejects_dot_segments_escaping_configured_path(requested: str):
+    """A requested path that resolves outside the configured path is not a hierarchical match."""
+    assert check_resource_allowed(requested, "https://example.com/api") is False
+
+
+def test_check_resource_allowed_resolves_dot_segments_on_both_sides():
+    """Both URLs are compared in resolved form, so equivalent spellings agree (SDK-defined matching)."""
+    assert check_resource_allowed("https://example.com/api/./v1", "https://example.com/api") is True
+    assert check_resource_allowed("https://example.com/api/v1", "https://example.com/other/../api") is True
+    assert check_resource_allowed("https://example.com/api/v1", "https://example.com/api/v1/../v2") is False
+
+
+def test_check_resource_allowed_keeps_encoded_slash_and_params_in_segment():
+    """`%2F` and `;params` are part of a segment (RFC 3986 sections 2.2, 3.3), not a boundary."""
+    assert check_resource_allowed("https://example.com/api%2Fv1", "https://example.com/api") is False
+    assert check_resource_allowed("https://example.com/api/v1", "https://example.com/api;x") is False