| Network Working Group | H. Tschofenig |
| Internet-Draft | E. Lear |
| Intended status: Informational | IAB Security Program |
| Expires: May 23, 2014 | November 19, 2013 |
Evolving the Web Public Key Infrastructure
draft-tschofenig-iab-webpki-evolution-01.txt
The problems with the WebPKI have received the attention by the Internet security community when DigiNotar, a Dutch certification authority, had a security breach in 2011 and in the same year a Comodo affiliate was compromised. Both cases led to fraudulent issuance of certificates and raise questions regarding the strength of the WebPKI used by so many applications.
Almost 2 years have passed since these incidents and various standardization activities have happened in the meanwhile offering new technical solutions to make the public key infrastructure more resilient.
The important question, however, is which of the technical solutions will get widespread deployment? In this document we compare the different technical solutions in an attempt to engage the impacted stakeholders to trigger deployment actions to improve the status quo. This document does not include any recommendations what techniques to use.
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at http://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on May 23, 2014.
Copyright (c) 2013 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
High-profile data breaches and security incidents on the Web are gaining increasing attention from the public, the press, and governments. A few examples may illustrate the problems: DigiNotar, a Dutch certification authority, had a security breach [DigiNotar] and in the same year a Comodo affiliate was compromised [Comodo]. Both cases led to fraudulent issuance of certificates.
The Web Public Key Infrastructure (WebPKI) makes use of trusted third parties, the certification authority (CAs), to bind a subject name to a public key. A CA may, however, be compromised despite the best security practices and operational procedures. The main problem, however, is that any CA can issue a certificate for any domain name. One compromised CA is therefore able to impact the security of the entire public key infrastructure. In the case of DigiNotar the attacker was able to issue certificates for Google services even though Google never made use of services from DigiNotar.
Furthermore, over time browser vendors and operating systems increased the number of trust anchors (TAs) that are pre-installed. Depending on software the number of trust anchors may exceed 600, as reported by the Electronic Frontier Foundation (EFF) in their SSL Observatory study [SSL-Observatory]. While the larger number provides choice for subjects regarding the CA they can select for obtaining a certificate there is also a downside: with today's WebPKI set-up it is sufficient to compromise a single CA to impact the security for all relying parties. Many users were surprised about the large number of trust anchors installed in normal operating systems and browsers without having an easy way to adjust that list to their preferences. Worse, most end users would not know what criteria to use to manage the trust anchor stores, even with better interfaces.
To re-state the problem statement: Every CA can issue certificates for any relying party even though that relying party may have never been in a relationship with the issuing CA. (Note that the trust anchor of that CA needs to be provisioned into the trust anchor store.)
These developments have led to a number of protocol design activities for improving the WebPKI. In this document we briefly summarize the available technical solutions and include an assessment about who needs to make changes, what type of benefits are provided, and what dependencies exist. The investigated solutions include DANE [RFC6698], Certificate Transparency [RFC6962], Public Key Pinning [I-D.ietf-websec-key-pinning], TACK [I-D.perrin-tls-tack], Perspectives [Perspectives], Sovereign Keys [SovereignKeys], MECAI [MECAI], Convergence [Convergence], and DetecTor [DetecTor].
There are many challenges with security on the Web, including user interface problems with certificate warnings, insecure use of cookies, cross-site scripting attacks, and injection attacks. This document focuses only on improving the WebPKI. It is also worth reminding ourselves that the WebPKI is not only used for Web applications but also for a range of other applications, including smart phone apps.
The main purpose of this document is to provide an overview of some technical solutions. This description will help us to develop a roadmap for the deployment of the best solutions to improve the overall security of the public key infrastructure.
Final note: There are also process solutions, such as stricter audits of CAs with the aim to improve operational practices, and these are not described in this document. These measures will be useful in addition to technical solutions but alone they will, however, not address the underlying problem.
This document uses the following terms from from RFC 5280 [RFC5280]:
This document also re-uses the term "Leap of faith" from RFC 5386 [RFC5386]:
RFC 6973 [RFC6973] provides a definition of the term 'relying party':
The terms 'trust anchor' and 'trust anchor store' are defined in [RFC6024]:
[I-D.ietf-websec-key-pinning] describes a solution for instructing user agents (UAs) to remember ("pin") certificates (end entity certificates or CA certs) for a given period of time. This pin is provided to the client with the initial interaction with a Web server using a newly defined HTTP headers. During that time, UAs will require that the TLS server presents a certificate chain including at least one Subject Public Key Info structure the fingerprint of which matches one of the pinned fingerprints for that server.
While the specification provides a number of instructions for the Website operator the basic operation is rather simple and assumes a leap-of-faith policy. To deal with the change of certificates or other failure scenarios, the concept of a backup pin is utilized. A Backup Pin is a fingerprint for the public key of a secondary, not-yet-deployed key pair. The operator keeps the backup key pair offline, and sets a pin for it in the Public-Key-Pins header. If an operator loses control of his/her primary private key, he/she can deploy the backup key pair. An interesting feature of the specification is to report pin validation failure to a URI using an HTTP POST.
When a pin validation failure occurs the expectation is that the user is notified about the inconsistency (with optionally reporting taking place in the background).
This document is the product of the IETF Web Security working group.
Similarly to the key pinning solution described in Section 3.1 TACK [I-D.perrin-tls-tack], also aims to enables a TLS server to support "pinning" to a self-chosen signing key.
A TLS server operator creates a so-called "TACK signing key" (TSK) and signs its own keys using this TSK. A TACK pin then associates a hostname, a TSK, and various parameters (including pin creation time, and lifetime of the pin). A TLS server operator may change a key for a server at any point in time since the TSK will be unchanged. Each server thereby acts as its own CA and issues end entity certificates to itself. Clients store the TACK pins in their pin stores, which they obtain via a TLS extension from the server directly. When TACK pins are obtained from the TLS server directly, they follow a leap-of-faith approach.
To enable incremental deployment, a TLS client uses the extension mechanism of TLS to indicate support for the TACK extension by including a new TLS extension type in the ClientHello message. A TLS server that does not support TACK will reply with an ordinary certificate. In case the TLS server supports the extension it replies with the newly defined tack structure, which contains the TACK pin for that server.
This specification is an individual submission to the IETF.
Perspectives [Perspectives] aims to utilize notaries (i.e., public servers) that monitor and record the history of public keys used by services. A notary cryptographically signs statements saying that at time t it observed service S using public key K. While the description focuses on the use of raw public keys (in the style of SSH) the same concept also works with certificates.
The basic approach is simple: When a TLS client starts to interact with a TLS server it is presented with a key/certificate that it had not seen before. To verify that the key/certificate is the same as observed by other vantage points in the network it contacts notary servers. These notaries provide key/certificate information they have obtained about the specific website earlier. Subsequently, clients cache the information returned by notaries for future use.
Since the security and availability of the proposed system depends on the location, the independence, and the number of notaries the following approach is suggested. Notaries are organized in groups, each group having a notary authority. The notary authority publishes lists of the notaries it offers. These lists contain entries with IP address information and their public keys, one for each notary. This list is then digitally signed with the private key of the notary authority. Clients are expected to download these lists and verify them with the public key of the respective notary authority. The public keys of the notary authorities are obtained out-of-band.
To improve the leap of faith security by clients, the notary services adds security value since they may have obtained the key/certificate from the website in the past already and from different vantage points (in terms of the path used to talk to the server). This helps when attacks are either temporary and or when a man-in-the-middle attacker is located somewhere along the path between the client and the server but closer to the client. The use of multiple notaries also helps to detect malicious notaries.
Similar to other notary services there is the question about how they obtain the keys/certificates of TLS servers (and other services). For popular services the keys/certificates are assumed to pre-configured at the notaries and queried periodically. For the long tail of small websites the suggested approach is to query these sites the first time a client wants to connect to them since this is the time when the notary learns about their existence.
With clients caching information about the keys/certificates of sites visited earlier, and the information obtained from notaries, there is no additional protocol overhead. In this respect the solution works similar to key pinning. The additional communication overhead for the client only occurs at the time when the client talks to a server for the first time or when the cached information expires.
The proposal is documented in form of an academic research paper, see [Perspectives], but no technical specification is available.
Convergence [Convergence] is a proposal by Moxie Marlinspike that makes two improvements to Perspectives, namely
As a main advantage, according to the author and similar to Perspectives, there is no impact on TLS server deployment, except in rare situations where multiple certificates are used by a single site in combination with a load balancer. Load balancers often terminate TLS connections and if those load balancers use independent certificates then different notaries requesting certificates from the Web site will receive different certificates.
Notaries are designed to be extensible by supporting different mechanisms with respect to how they obtain certificates. Currently, Convergence uses the technique proposed by Perspectives to probe a TLS server.
The documentation of Convergence only exists in form of a presentation by Moxie Marlinspike given at the BlackHat USA 2011 conference [ConvergenceTalk].
Sovereign Keys [SovereignKeys] is a proposal by the Electronic Frontier Foundation (EFF). It introduces two new concepts to deal with attacks against the WebPKI.
Each timeline server possesses a unique private/public key pair and these keys are assumed to be shipped with client software or TLS libraries to ensure that clients can verify the authenticity of timeline entries. The timeline servers record the history of claims to sovereign keys.
TLS clients query timeline servers for entries that belong to a certain domain and verify that the end-entity certificate has been issued under the sovereign key. If the verification fails then the connection attempt is refused.
A high-level description can be found at [SovereignKeys] and a more detailed technical specification is available at [SovereignKeys-Spec].
MECAI [MECAI] builds conceptually on top of the Perspective proposal. Perspectives introduces notaries, as new entities in the WebPKI, and MECAI takes the position that this function can be taken by existing CAs. With this new role they would turn into Voucher Authorities (VAs), who issue vouchers that confirm what they observe. A voucher is a signature computed over a number of fields including the hash of a server certificate, a certificate chain, the IP address of the server, revocation status information, etc. Of course, a voucher would be created by a CA other than the one that created the original certificate.
A client would therefore perform the following steps: it connects to a server via TLS and the server provides the certificate. Then, the client needs to obtain one or multiple vouchers for the server certificate. This can happen either inband within the TLS handshake when talking to the server, similarly to how OCSP stapling works, or via a separate protocol exchange. The former approach is less expensive in terms of communication costs for the client. In any case, a voucher request protocol is needed to let entities (like TLS servers) talk to VAs to obtain a voucher.
A client or a server can detect misissuance by matching the information in the vouchers with the certificate.
Only a high-level description is available via [MECAI] but no detailed technical specification.
DANE [RFC6698] offers the option to use the DNS infrastructure to store certificates. DANE is envisioned as a preferable basis for binding public keys to DNS names, because the entities that vouch for the binding of public key data to DNS names are the same entities responsible for managing the DNS names in question.
Distributing certificates via the DNS does, however, require DNSSEC. With the help of DNSSEC [RFC4033][RFC4034][RFC4035] this offers an opportunity to eliminate off-line processes for validation of the subject name, which today often requires sending a mail to the administrator of that domain. This relationship can be easily demonstrated by having the zone administrator for the subject domain post the public key in the DNS and digitally sign the resulting zone.
A high-level description about the different options offered by DANE can be found in [IETF-Journal-DANE] and the authoritative version can be found in RFC 6698 [RFC6698].
RFC 6962 [RFC6962] specifies Certificate Transparency, a protocol for publicly logging the existence of certificates as they are issued or observed. This allows anyone to audit certification authority (CA) activity and detect the issuance of suspect certificates, as well as to audit the certificate logs themselves. The intent is that eventually clients would refuse to honor certificates that do not appear in a log, effectively forcing CAs to add all issued certificates to the logs.
The publicly auditable, append-only logs of all issued certificates does not prevent misissue but allows interested parties to detect misissuance.
While various projects, including the EFF with their SSL Observatory [SSL-Observatory] and Crossbear [Crossbear], have scanned the Internet to collect all certificates of publically accessible TLS servers, the cooperation from all TAs (and subordinate CAs) is required to make Certificate Transparency proposal successful. The reasons are two-fold: IPv6 makes scanning the address range of the entire Internet much more difficult and the increasing deployment of the TLS Server Name Indication [RFC6066] prevents it from obtaining all certificates.
The expected operation is as follows: TA and subordinate CAs contact log servers and upload certificates, as they are issued. In response, they receive a Signed Certificate Timestamp (SCT). The SCT is the log's promise to incorporate the certificate in a Merkle Tree, which is the data structure used by the log, within a fixed amount of time. Everyone can check the log for consistency. Website operators will have an incentive to regularly check the logs for misissuance of certificates issued to DNS names associated with their sites. TLS clients on the other hand are not expected to directly communicate with logs to avoid the communication overhead. Instead, the TLS servers provides the SCT along with the certificate within the TLS handshake. TLS clients reject certificates that do not have a valid SCT for the end entity certificate. Since there is ideally more than one log TLS servers need to provide SCTs from multiple logs to the client.
This document has gone through a public review process, and has been approved by the Internet Engineering Steering Group and published an experimental RFC.
DetecTor [DetecTor] extends the idea of MECAI and Perspectives by utilizing the Tor onion routing infrastructure [Tor] in order to connect to sites via different paths through the network. The Tor infrastructure thereby replaces the need to have dedicated notary servers, who connect to sites to obtain certificates from a different vantage point. The server certificate obtained via one or multiple Tor connections is then compared with the certificate that was obtained via the direct TLS connection between the client and the site (i.e., without using Tor). This offers capabilities for the client to detect whether there was an adversary along the path but close to the client.
Unlike other proposals, the suggestion is made to provide no information to the user once a failure has been detected. Instead, the connection attempt will be rejected and no recourse is possible. Like other proposals information about the observed certificates may be cached by the client to lower the initial set-up delay.
A high-level description can be found at [DetecTor] but no detailed technical specification is available.
This version of the document re-uses the analysis criteria proposed by Eric Rescorla [Rescorla].
A common problem of all proposals that aim to prevent attacks lies in the user interface design when a failure occurs and end users are informed about the problem. In many cases, the failure may not necessarily be caused by real attacks but rather by the use of captive portals or server-side configuration problems (like warnings caused by expired certificates today). User interface studies, such as [SE09], [SR07], and [BO09], have shown that end users are typically not in the best position to make judgments about these security warning dialogs. Furthermore, proposals that make use of out-of-band communication interactions may face problems with firewalled networks and the additional delay incurred. Claims have been made that this is a problem with the use of OCSP today [OCSP-Performance], which has been the motivation for developing and standardizing OCSP stapling and multiple OCSP stapling.
This entire document is about security.
The main privacy threat is correlation. Correlation is the combination of various pieces of information related to an individual or that obtain that characteristic when combined. In this specific case there is the risk that newly introduced entities obtain information about the history of service usage by users. For example, a notary that is contacted each time a user visits a new website can easily be seen as problematic from a privacy point of view.
This document does not require actions by IANA.
We would like to thank all participants of the NIST workshop on "Improving Trust in the Online Marketplace", April 10-11 2013, for sharing their views with the community. We would also like to thank the authors of various solution proposals for their work.
Big thanks to Steven Kent for his detailed review. Additionally, we would like to thank Ben Laurie for his feedback.