title: "Reputation Verified Selection of Upstream Encrypted Resolvers" abbrev: "DDR and Forwarders" docname: draft-schwartz-add-ddr-forwarders-latest category: info
ipr: trust200902 area: General workgroup: add keyword: Internet-Draft
stand_alone: yes pi: [toc, sortrefs, symrefs]
ins: B. Schwartz
name: Benjamin Schwartz
organization: Google LLC
email: [email protected]
- ins: C. Box name: Chris Box organization: BT email: [email protected]
informative: FIREFOX-FALLBACK: target: https://support.mozilla.org/en-US/kb/firefox-dns-over-https#w_about-our-rollout-of-dns-over-https title: About our rollout of DNS over HTTPS MOZILLA-CANARY: target: https://support.mozilla.org/en-US/kb/canary-domain-use-application-dnsnet title: Canary domain - use-application-dns.net DUCK-CNAME: target: https://help.duckduckgo.com/duckduckgo-help-pages/features/safe-search/ title: Force Safe Search at a Network Level BING-CNAME: target: https://help.bing.microsoft.com/#apex/bing/en-us/10003/0 title: Block adult content with SafeSearch - Map at a network level GOOGLE-CNAME: target: https://support.google.com/websearch/answer/186669?hl=en title: Keep SafeSearch turned on for your school, workplace, or home network MOZILLA-TRR: target: https://wiki.mozilla.org/Security/DOH-resolver-policy#Mozilla_Policy_Requirements_for_DNS_over_HTTPs_Partners title: Mozilla Policy Requirements for DNS over HTTPs Partners CHROME-DOH: target: https://docs.google.com/document/d/128i2YTV2C7T6Gr3I-81zlQ-_Lprnsp24qzy_20Z1Psw/edit title: "DoH providers: criteria, process for Chrome" MICROSOFT-DOH: target: https://docs.microsoft.com/en-us/windows-server/networking/dns/doh-client-support#determine-which-doh-servers-are-on-the-known-server-list title: Determine which DoH servers are on the known server list NEXTDNS: target: https://nextdns.io/ title: NextDNS
--- abstract
This draft describes an extension to the Discovery of Designated Resolvers (DDR) standard, enabling use of encrypted DNS in the presence of legacy DNS forwarders.
--- middle
The Discovery of Designated Resolvers specification {{?DDR=I-D.draft-ietf-add-ddr}} describes a mechanism for clients to learn about the encrypted protocols supported by a DNS server. It also describes a client validation policy that has strong security properties.
Recent estimates suggest that a large fraction, perhaps a majority, of residential internet users in the United States and Europe rely on local DNS forwarders that are not compatible with DDR. This is because they are accessed via a private IP address, which TLS certificates cannot normally prove ownership of. Many such devices also face significant hurdles in being upgraded to support encrypted DNS, so it is likely that a large installed base of legacy DNS forwarders, providing Do53 on a private IP address, will remain for some years.
A client in such a network that wants to use the network's DNS resolver is forced to use Do53. It is therefore vulnerable to passive surveillance both on the local network, and between this network and the upstream provider, even if the upstream DNS resolver supports encrypted DNS.
Many of these attacks can be mitigated by using the method described in this document. In a nutshell the process is as follows.
- The client begins DDR discovery, querying for _dns.resolver.arpa.
- The legacy DNS forwarder, since it does not understand DDR, forwards this query upstream.
- The upstream recursive resolver, which supports DDR, replies with details of how to access its encrypted DNS service.
- The client receives this response and performs Reputation Verified Selection (see {{client-policy}}).
- On successful completion, the client may commence using encrypted DNS towards the upstream resolver. This is known as Cross-Forwarder Upgrade.
By this process, Do53 is replaced with encrypted DNS for most queries. The client may wish to continue to send locally-relevant queries (e.g. .local) towards the legacy DNS forwarder.
This document describes the interaction between DDR and legacy DNS forwarders.
DNS forwarders and resolvers that are implemented with awareness of DDR are out of scope, as they are not affected by this discussion (although see Security Considerations, {{security-considerations}}).
IPv6-only networks whose default DNS server has a Global Unicast Address are out of scope, even if this server is actually a simple forwarder. If the DNS server does not use a private IP address, it is not a "legacy DNS forwarder" under this draft's definition.
Private IP Address - Any IP address reserved for loopback {{?RFC1122}}, link-local {{?RFC3927}}, private {{?RFC1918}}, local {{?RFC4193}}, or Carrier-Grade NAT {{?RFC6598}} use.
Legacy DNS Forwarder - An apparent DNS resolver, known to the client only by a private IP address, that forwards the client's queries to an upstream resolver, and has not been updated with any knowledge of DDR.
Cross-Forwarder Upgrade - Establishment and use of a direct, encrypted connection between the client and the upstream resolver.
Reputation Verified Selection (RVS) is a method for validating whether connection using DDR is allowed. Clients MAY use RVS when (a) the local DNS server is identified by a Private IP address and (b) the DDR SVCB resolution process does not produce any Encrypted DNS endpoints that have this IP address in their A or AAAA records. RVS then proceeds as follows:
- The client connects to one of the indicated Encrypted DNS endpoints.
- The client receives a certificate, which it verifies to a suitable root of trust.
- For each identity (e.g. SubjectAltName) in the certificate, the client constructs a Resolver Identity:
a. For DNS over TLS and DNS over QUIC, the Resolver Identity is an IP address or hostname and the port number used for the connection.
b. For DNS over HTTPS, the Resolver Identity is a URI Template in absolute form, containing the port number used for the connection and path indicated by
dohpath. - The client determines the reputation of each Resolver Identity derived from the certificate.
- The maximum (i.e. most favorable) reputation is the reputation of this connection.
OPEN QUESTION: Would it be better to use the SVCB TargetName to select a single Resolver Identity? This would avoid the need to enumerate the certificate's names, but it would require the use of SNI (unlike standard DDR), and would not be compatible with all upstream encrypted resolvers.
OPEN QUESTION: Can we simplify the resolver identity to just a domain name? This would make reputation systems easier, but it would not allow distinct reputation for different colocated resolution services, so reputation providers would have to be sure that no approved resolver has other interesting colocated services.
This process MUST be repeated whenever a new TLS session is established, but reputation scores for each resolver endpoint MAY be cached.
For DNS over HTTPS, the :authority pseudo-header MUST reflect the Resolver Identity with the most favorable reputation, to ensure that the HTTP requests are well-formed and are directed to the intended service. If the Resolver Identity is a wildcard, the reputation system MUST replace it with a valid hostname that matches the wildcard.
Assessing reputation limits the ability of a DDR forgery attack to cause harm, as it will only allow an attacker to direct clients to a resolver they consider trustworthy. Major DoH client implementations already include lists of known or trusted resolvers {{CHROME-DOH}}{{MICROSOFT-DOH}}{{MOZILLA-TRR}}.
If no resolvers pass the reputation check, the client must not proceed.
Clients SHOULD start by checking the resolver endpoint with the numerically lowest SVCB SvcPriority. Clients MAY wait until a DNS query triggers an Encrypted DNS connection attempt before performing this verification.
If RVS encounters an error or rejects the server, the client MUST fall back to plaintext DNS on port 53.
Successful validation then permits cross-forwarder upgrade.
Embedding a list of known trusted resolvers in a client is only one possible model for assessing the reputation of a resolver. In future a range of online reputation services might be available to be queried, each returning an answer according to their own specific criteria. These might involve answers on other properties such as jurisdiction, or certification by a particular body. It is out of scope for this document to define these query methods, other than to note that designers should be aware of bootstrapping problems. It is the client's decision as to how to combine these answers, possibly using additional metadata (e.g. location), to make a determination of reputation.
If the determined reputation is a binary "definitely trustworthy" or "definitely malicious", the client's recommended action is clear. However, intermediate trust levels are also possible (e.g. "probably safe", "newly launched"). In these cases there are some options clients can consider.
The client can simply decline to the use the encrypted service. In this case, unless there is another option, the client will fall back to Do53.
The client can ask the user about specific domain names that appear in the certificate. These names might be recognizable to the user, e.g. as that of an ISP. It's also possible to convey information about why the ADN lacks some element of reputation.
The client can also use the encrypted service for a limited time, as a means of mitigating. By limiting the DDR response TTL to 5 minutes, a client can ensure that any attacker can continue to monitor queries for at most 5 minutes after they have left the local network.
Certain local DNS forwarders block access to domains associated with malware and other threats. Others block based on the category of service provided by those domains, e.g. domains hosting services that are not appropriate for a work or school environment. In the short term to ensure this service is not lost due to a cross-forwarder upgrade, the maintainers can simply add "resolver.arpa" to their actively curated list of domains to block. This pattern has been deployed by Mozilla, with the domain "use-application-dns.net" {{MOZILLA-CANARY}}.
In the long term, it is best for filtering DNS forwarders to implement support for encrypted DNS. The following subsections describe some ways to implement this.
The local forwarder can be upgraded to one that implements an encrypted DNS service discoverable through DNR. This requires a TLS certificate on the local device, proving ownership of the chosen Authentication Domain Name (ADN). Onward queries to the internet SHOULD also be protected with encryption.
If the local forwarder can be upgraded to offer an encrypted DNS service, this can then be made discoverable through classic DDR. If the device has a private IP (as presumed for RVS), a self-signed certificate is sufficient as long as the client supports the Opportunistic Discovery mode of DDR. Onward queries to the internet SHOULD also be protected with encryption.
The blocking functionality can be moved to the upstream resolver. Cross-forwarder upgrade then enables the service to continue, as long as the upstream resolver has sufficient reputation.
There are a small number of other issues to be aware of. For all these, a possible general mitigation is to provide users or administrators with the ability to control whether DDR is used with legacy forwarders. For example, this control could be provided via a preference, or via a notification upon discovering a new upstream resolver. Specific mitigations are also described below.
Some local network resolvers contain additional names that are not resolvable in the global DNS. A simple cross-forwarder upgrade might lose access to these local names. Clients should be aware of well-known suffixes (e.g. .local, .home.arpa.) that require local resolution. Dynamic discovery of local prefixes would help this issue. To address any remaining ones, the following mitigation can be used.
In "NXDOMAIN Fallback", the client repeats a query to the unencrypted resolver if the encrypted resolver returns NXDOMAIN. This allows the resolution of local names, provided they do not collide with globally resolvable names (as required by {{?RFC2826}}).
This is similar to the fallback behavior currently deployed in Mozilla Firefox {{FIREFOX-FALLBACK}}.
NXDOMAIN Fallback results in slight changes to the security and privacy properties of encrypted DNS. Queries for nonexistent names no longer have protection against a local passive adversary, and local names are revealed to the upstream resolver.
NXDOMAIN Fallback is only applicable when a legacy DNS forwarder might be present, i.e. the unencrypted resolver has a private IP address, and the encrypted resolver has a different IP address. In other DDR configurations, any local names are expected to resolve similarly on both resolvers.
An "interposable domain" is a domain whose owner deliberately allows resolvers to forge certain responses. This arrangement is most common for search engines, which often support a configuration where resolvers forge a CNAME record to direct all clients to a child-appropriate instance of the search engine {{DUCK-CNAME}}{{BING-CNAME}}{{GOOGLE-CNAME}}.
Future deployments of interposable domains can instruct administrators to enable or disable DDR when adding the forged record, but forged records in legacy DNS forwarders could be lost due to a cross-forwarder upgrade.
There are a small number of pre-existing interposable domains, largely of interest only to web browsers. Clients can maintain a list of relevant interposable domains and resolve them only via the network's resolver.
Many legacy DNS forwarders also provide a shared cache for all network users. Cross-forwarder upgrades will bypass this cache, resulting in slower DNS resolution for some queries.
Clients can compensate partially for any loss of shared caching by implementing local DNS caches. This mitigation is already widely deployed in browsers and operating systems.
The conservative validation policy results in no encryption when a legacy DNS forwarder is present. This leaves the user's query activity vulnerable to passive monitoring {{?RFC7258}}, either on the local network or between the user and the upstream resolver.
Reputation Validated Selection enables the use of encrypted transport in these configurations, reducing exposure to a passive surveillance adversary.
When the client uses the conservative validation policy described in {{DDR}}, the client can establish a secure DDR connection only in the absence of an active attacker. An on-path attacker can impersonate the resolver and intercept all queries, by preventing the DDR upgrade.
This basic security property also applies if the client uses reputation validated selection, but an additional one is added.
An on-path attacker might be located on the local network, or between the local network and the upstream resolver. In either case, the attacker can redirect the client to a resolver of the attacker's choice, /as long as that resolver meets the client's requirements for reputation/. Hence the reputation system is essential to the security of the user. If a previously-reputable resolver is compromised, users can be redirected to it while this reputation remains high.
Once an attack has been detected, it should be reported to relevant reputation services so that they can revise their assessment of the ADN.
With the conservative validation policy, a random sample of IP packets is likely sufficient for manual retrospective detection of an active attack.
With reputation verified selection, forensic logs must capture a specific packet (the attacker’s DDR designation response) to enable retrospective detection.
Network-layer forensic logs that are not integrated with the resolver can enable detection of these attacks by logging all DDR responses, or more generally all DNS responses. This makes retrospective attack detection straightforward, as the attacker's DDR response will indicate an unexpected server.
DNS-layer forensic logging conducted by a legacy DNS forwarder would be lost in a cross-forwarder upgrade.
Forwarders that want to observe all queries from RVS clients should plan to implement DDR or DNR. In the short term it is possible for the forwarder to disable DDR by responding negatively to _dns.resolver.arpa but this is not recommended long-term as it prevents confidentiality protection.
With Do53 to a legacy DNS forwarder, an on-path attacker located between the local network and the upstream resolver is not directly aware of how many devices are making DNS queries behind the forwarder. It can only see aggregated queries being made by the forwarder. {{DDR}} to a non-local resolver permits the attacker to become aware of the individual encrypted DNS connections from each device, noting how many there are and the relative number of queries/responses made for each. RVS shares this property.
If the above issue is a concern, clients may wish to open a random number of connections to the designed encrypted resolver and distribute queries among them. This may lead the attacker to assume a larger number of devices than are actually present.
--- back
{:numbered="false"}
Thanks to Anthony Lieuallen and Eric Orth for early reviews of a previous draft.