fix(noema): fail closed at the credential egress boundary

CHANGELOG.md

@@ -70,6 +70,16 @@ Semantic Versioning where the repository publishes a release.
 
 ### Fixed
 
+- Pin Noema's HTTP and HTTPS credential egress to the prevalidated numeric DNS
+  addresses while preserving the original HTTPS hostname for certificate
+  validation, preventing DNS rebinding between endpoint validation and socket
+  creation.
+- Fail closed before sending Noema's bearer credential to a configured model
+  endpoint unless a non-loopback target uses HTTPS with stable, globally
+  routable unicast DNS evidence; preserve only the literal same-job loopback
+  sidecar exception, keep redirects disabled, and bound response bodies to
+  1 MiB before JSON decoding.
+
 - Publish only the sanitized cumulative Strix report tree, avoiding a later
   copy of relative scanner output that could reintroduce known internal warning
   text into uploaded security evidence.

docs/doctoring/noema-credential-egress-boundary.md

@@ -0,0 +1,57 @@
+# Noema credential-egress boundary
+
+## Customer outcome
+
+Noema now rejects a misconfigured or attacker-influenced model endpoint before
+placing its bearer credential on the request. A public service must use HTTPS,
+must resolve entirely to globally routable unicast addresses, and must retain
+the same complete DNS address set after the bounded response is received.
+Resolver failures, empty answers, malformed addresses, redirects, special-use
+addresses, and response bodies larger than 1 MiB fail closed.
+
+The existing same-job contextual-orchestrator seam remains usable without
+importing the orchestrator into this repository: literal `127.0.0.1` and `::1`
+endpoints may use HTTP only when every resolver result is loopback. Hostnames
+that merely resolve to loopback do not receive this exception. Provider routing,
+model selection, and model-parameter translation remain upstream concerns.
+
+## Decision and trust boundary
+
+The implementation reuses Python's `urllib.parse`, `socket.getaddrinfo`, and
+`ipaddress` rather than adding a URL or address-classification dependency.
+Before request construction it:
+
+1. rejects non-HTTP schemes, URL user information, and missing hostnames;
+2. restricts plaintext HTTP to the two literal loopback sidecar addresses;
+3. resolves the endpoint for its effective port and rejects failed, empty, or
+   malformed resolution evidence; and
+4. requires every non-loopback address to be globally routable unicast.
+
+Redirects remain disabled, and the opener disables ambient proxies. Custom HTTP
+and HTTPS connections connect only to the validated numeric addresses while
+retaining the original hostname for HTTPS certificate validation. The response
+reader requests at most one byte beyond the 1 MiB contract, rejects an
+over-limit result before decoding JSON, and then re-resolves the same host and
+port. A changed address set invalidates the result. Tests cover IPv4 and IPv6
+loopback, dual-stack public endpoints, pinned numeric TCP destinations, TLS SNI,
+DNS rebinding evidence, resolver failures, malformed results, URL credentials,
+special-use address classes, and the byte limit.
+
+Trusted DNS and network egress controls remain defense-in-depth for production;
+the application boundary now also prevents the request socket from performing
+an unvalidated hostname resolution.
+
+## References
+
+Cotton, M., Vegoda, L., Bonica, R., & Haberman, B. (2013). *Special-purpose IP
+address registries* (RFC 6890). RFC Editor. https://doi.org/10.17487/RFC6890
+
+MITRE. (2026a). *CWE-400: Uncontrolled resource consumption*.
+https://cwe.mitre.org/data/definitions/400.html
+
+MITRE. (2026b). *CWE-918: Server-side request forgery (SSRF)*.
+https://cwe.mitre.org/data/definitions/918.html
+
+OWASP Foundation. (n.d.). *Server-side request forgery prevention cheat sheet*.
+Retrieved August 24, 2026, from
+https://cheatsheetseries.owasp.org/cheatsheets/Server_Side_Request_Forgery_Prevention_Cheat_Sheet.html

scripts/ci/noema_review_gate.py

@@ -5,6 +5,7 @@
 
 import argparse
 import base64
+import http.client
 import ipaddress
 import json
 import os
@@ -41,6 +42,11 @@
 MAX_FILE_CONTEXT_CHARS = 4000
 MAX_REVIEW_CONTEXT_CHARS = 24000
 MAX_THREAD_BODY_CHARS = 1200
+MAX_LLM_RESPONSE_BYTES = 1024 * 1024
+IpAddress = ipaddress.IPv4Address | ipaddress.IPv6Address
+TRUSTED_LOOPBACK_ADDRESSES = frozenset(
+    {ipaddress.ip_address("127.0.0.1"), ipaddress.ip_address("::1")}
+)
 
 # ⚡ Bolt: Pre-compiled regex patterns to avoid recompilation on every scrub_sensitive_data call.
 # Impact: Improves string processing performance in error reporting.
@@ -423,6 +429,172 @@ def redirect_request(
         raise urllib.error.HTTPError(req.full_url, code, msg, headers, fp)
 
 
+def _socket_target(address: IpAddress, port: int) -> tuple[str, int]:
+    """Return a socket destination that contains only a validated IP literal."""
+    return (str(address), port)
+
+
+class _PinnedConnectionMixin:
+    """Connect only to the DNS addresses validated before the request."""
+
+    def __init__(
+        self,
+        *args: Any,
+        validated_addresses: frozenset[IpAddress],
+        **kwargs: Any,
+    ) -> None:
+        """Store immutable DNS evidence before initializing the HTTP connection."""
+        if not validated_addresses:
+            raise ValueError("Noema endpoint has no validated DNS addresses")
+        self._validated_addresses = tuple(
+            sorted(validated_addresses, key=lambda address: (address.version, address.packed))
+        )
+        super().__init__(*args, **kwargs)
+
+    def _connect_to_validated_address(self) -> None:
+        """Open a socket using a validated numeric destination, never the hostname."""
+        last_error: OSError | None = None
+        for address in self._validated_addresses:
+            try:
+                self.sock = socket.create_connection(
+                    _socket_target(address, self.port),
+                    self.timeout,
+                    self.source_address,
+                )
+                return
+            except OSError as exc:
+                last_error = exc
+        raise OSError("Noema endpoint could not connect to validated DNS addresses") from last_error
+
+
+class PinnedHTTPConnection(_PinnedConnectionMixin, http.client.HTTPConnection):
+    """HTTP connection that cannot resolve the configured hostname a second time."""
+
+    def connect(self) -> None:
+        """Connect to a validated address and preserve proxy tunnel behavior."""
+        self._connect_to_validated_address()
+        if self._tunnel_host:
+            self._tunnel()
+
+
+class PinnedHTTPSConnection(_PinnedConnectionMixin, http.client.HTTPSConnection):
+    """HTTPS connection pinned to validated addresses while retaining hostname TLS."""
+
+    def connect(self) -> None:
+        """Connect to a validated address, then verify the original hostname in TLS."""
+        self._connect_to_validated_address()
+        if self._tunnel_host:
+            self._tunnel()
+        server_hostname = self._tunnel_host or self.host
+        self.sock = self._context.wrap_socket(self.sock, server_hostname=server_hostname)
+
+
+class PinnedHTTPHandler(urllib.request.HTTPHandler):
+    """urllib handler that uses validated numeric destinations for HTTP requests."""
+
+    def __init__(self, addresses: frozenset[IpAddress]) -> None:
+        """Bind this handler to one prevalidated DNS result set."""
+        super().__init__()
+        self._addresses = addresses
+
+    def http_open(self, req: urllib.request.Request) -> Any:
+        """Open an HTTP request without resolving its hostname again."""
+        return self.do_open(
+            lambda host, **kwargs: PinnedHTTPConnection(
+                host, validated_addresses=self._addresses, **kwargs
+            ),
+            req,
+        )
+
+
+class PinnedHTTPSHandler(urllib.request.HTTPSHandler):
+    """urllib handler that pins TCP while preserving HTTPS hostname verification."""
+
+    def __init__(self, addresses: frozenset[IpAddress]) -> None:
+        """Bind this handler to one prevalidated DNS result set."""
+        super().__init__()
+        self._addresses = addresses
+
+    def https_open(self, req: urllib.request.Request) -> Any:
+        """Open HTTPS using the validated address set and original URL hostname."""
+        return self.do_open(
+            lambda host, **kwargs: PinnedHTTPSConnection(
+                host, validated_addresses=self._addresses, **kwargs
+            ),
+            req,
+            context=self._context,
+        )
+
+
+def resolve_endpoint_addresses(
+    hostname: str,
+    port: int,
+) -> frozenset[IpAddress]:
+    """Resolve every stream address for an endpoint or fail closed."""
+    try:
+        addrinfo = socket.getaddrinfo(hostname, port, type=socket.SOCK_STREAM)
+    except (OSError, UnicodeError) as exc:
+        raise ValueError("Noema endpoint DNS resolution failed") from exc
+    if not addrinfo:
+        raise ValueError("Noema endpoint DNS resolution returned no addresses")
+
+    addresses: set[IpAddress] = set()
+    for result in addrinfo:
+        try:
+            raw_address = result[4][0]
+            address = ipaddress.ip_address(raw_address)
+        except (IndexError, TypeError, ValueError) as exc:
+            raise ValueError("Noema endpoint DNS resolution returned a malformed address") from exc
+        addresses.add(address)
+    return frozenset(addresses)
+
+
+def validate_endpoint(
+    api_url: str,
+) -> tuple[str, int, frozenset[IpAddress]]:
+    """Validate transport and pre-request DNS evidence for a model endpoint."""
+    if not (api_url.lower().startswith("http://") or api_url.lower().startswith("https://")):
+        raise ValueError(
+            "URL scheme must be http or https; NOEMA_LLM_API_URL must start "
+            "with http:// or https:// to prevent SSRF vulnerabilities"
+        )
+    parsed = urllib.parse.urlparse(api_url)
+    scheme = parsed.scheme.lower()
+    if scheme not in {"http", "https"}:
+        raise ValueError(
+            "URL scheme must be http or https; NOEMA_LLM_API_URL must start "
+            "with http:// or https://"
+        )
+    hostname = (parsed.hostname or "").lower()
+    if not hostname:
+        raise ValueError("URL must have a valid hostname")
+    if parsed.username is not None or parsed.password is not None:
+        raise ValueError("Noema endpoint URL cannot contain user information")
+    if hostname in {"localhost", "localhost.localdomain"} or hostname.endswith(".localhost"):
+        raise ValueError("URL cannot target localhost")
+
+    port = parsed.port or (443 if scheme == "https" else 80)
+    try:
+        literal_address = ipaddress.ip_address(hostname)
+    except ValueError:
+        literal_address = None
+    trusted_loopback = literal_address in TRUSTED_LOOPBACK_ADDRESSES
+    if not trusted_loopback and scheme != "https":
+        raise ValueError("Noema non-loopback endpoints must use HTTPS")
+
+    addresses = resolve_endpoint_addresses(hostname, port)
+    if trusted_loopback:
+        if any(not address.is_loopback for address in addresses):
+            raise ValueError("Noema loopback endpoint DNS resolution left loopback")
+    elif any(
+        not address.is_global or address.is_multicast for address in addresses
+    ):
+        raise ValueError(
+            "Noema non-loopback endpoint DNS must contain only globally routable unicast addresses"
+        )
+    return hostname, port, addresses
+
+
 def extract_json_object(text: str) -> dict[str, Any]:
     """Extract a JSON object from a strict or lightly wrapped LLM response."""
     stripped = text.strip()
@@ -449,32 +621,7 @@ def call_llm(
     model = os.environ.get("NOEMA_LLM_MODEL", "").strip() or "noema-default"
     if not api_url or not api_key:
         raise RuntimeError("Noema LLM review unavailable: NOEMA_LLM_API_URL or NOEMA_LLM_API_KEY is not configured.")
-    if not (api_url.lower().startswith("http://") or api_url.lower().startswith("https://")):
-        raise ValueError(
-            "URL scheme must be http or https; NOEMA_LLM_API_URL must start "
-            "with http:// or https:// to prevent SSRF vulnerabilities"
-        )
-    parsed = urllib.parse.urlparse(api_url)
-    if parsed.scheme.lower() not in {"http", "https"}:
-        raise ValueError("URL scheme must be http or https; NOEMA_LLM_API_URL must start with http:// or https://")
-    hostname = (parsed.hostname or "").lower()
-    if not hostname:
-        raise ValueError("URL must have a valid hostname")
-    if hostname in {"localhost", "localhost.localdomain"} or hostname.endswith(".localhost"):
-        raise ValueError("URL cannot target localhost")
-    try:
-        addrinfo = socket.getaddrinfo(hostname, None)
-    except socket.gaierror:
-        pass
-    else:
-        for result in addrinfo:
-            ip_str = result[4][0]
-            try:
-                ip = ipaddress.ip_address(ip_str)
-            except ValueError:
-                continue
-            if ip.is_private or ip.is_loopback or ip.is_link_local or ip.is_multicast or ip.is_unspecified:
-                raise ValueError("URL cannot target internal IP addresses")
+    hostname, port, addresses = validate_endpoint(api_url)
 
     prompt = {
         "role": "user",
@@ -514,9 +661,21 @@ def call_llm(
         },
         method="POST",
     )
-    opener = urllib.request.build_opener(NoRedirectHandler())
+    opener = urllib.request.build_opener(
+        # Force the pinned direct path; proxy destinations have not passed this
+        # endpoint's DNS policy and must not receive the bearer credential.
+        urllib.request.ProxyHandler({}),
+        PinnedHTTPHandler(addresses),
+        PinnedHTTPSHandler(addresses),
+        NoRedirectHandler(),
+    )
     with opener.open(request, timeout=120) as response:  # nosec B310
-        raw = response.read().decode("utf-8")
+        raw_bytes = response.read(MAX_LLM_RESPONSE_BYTES + 1)
+    if len(raw_bytes) > MAX_LLM_RESPONSE_BYTES:
+        raise RuntimeError("Noema LLM response exceeded the byte limit")
+    if resolve_endpoint_addresses(hostname, port) != addresses:
+        raise ValueError("Noema endpoint DNS addresses changed during the request")
+    raw = raw_bytes.decode("utf-8")
     data = json.loads(raw)
     content = (((data.get("choices") or [{}])[0].get("message") or {}).get("content") or "").strip()
     verdict = extract_json_object(content)