| Internet Architecture Board | R. Housley |
| Internet-Draft | Vigil Security |
| Intended status: Informational | K. O'Donoghue |
| Expires: August 24, 2016 | Internet Society |
| February 21, 2016 |
Problems with the Public Key Infrastructure (PKI) for the World Wide Web
draft-iab-web-pki-problems-01.txt
This document describes some of the challenges facing the current Public Key Infrastructure (PKI) used for the World Wide Web (Web PKI) and considers promising improvements to address these challenges. Technical, process, and policy improvements to the WebPKI are considered. In addition, some technical considerations beyond WebPKI itself are considered. Hopefully the content of this document will help drive the Internet community toward wide spread adoption of some of the highlighted recommendations.
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 August 24, 2016.
Copyright (c) 2016 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.
There are many technical, process, and policy problems with the current Public Key Infrastructure (PKI) used for the World Wide Web (Web PKI). This document describes these problems, considers some emerging technical improvements, discusses some evoling process and policy improvements, and provides some basic recommendations for the Internet community.
The Web PKI makes use of certificates as described in RFC 5280 [RFC5280]. These certificates are primarily used with Transport Layer Security (TLS) RFC 5246 [RFC5246].
Certificates are specified in RFC 5280 [RFC5280]. Certificates contain, among other things, a subject name, a public key, a limited valid lifetime, and the digital signature of the Certification Authority (CA). Certificate users require confidence that the private key associated with the certified public key is owned by the named subject.
The architectural model used in the Web PKI includes:
A valid certificate may be revoked for any number of reasons. CAs are responsible for indicating the revocation status of the certificates that they issue throughout the lifetime of the certificate. Revocation status information may be provided using the Online Certificate Status Protocol (OCSP) RFC 6960 [RFC6960], certificate revocation lists (CRLs) RFC 5280 [RFC5280], or some other mechanism. In general, when revocation status information is provided using CRLs, the CA is also the CRL issuer. However, a CA may delegate the responsibility for issuing CRLs to a different entity.
The enrollment process used by a CA makes sure that the subject name in the certificate is appropriate and that the subject actually holds the private key. Proof of possession of the private key is often accomplished through a challenge-response protocol.
Over the years, many technical improvements have been made to the Web PKI. This section discusses several problems and the technical improvements that have been made to address them. This history sets the stage for suggestions for additional improvements in other sections of this document.
Over the years, the digital signature algorithms, one-way hash functions, and public key sizes that are considered strong have changed. This is not a surprise. Cryptographic algorithms age; they become weaker with time. As new cryptanalysis techniques are developed and computing capabilities improve, the work factor to break a particular cryptographic algorithm will reduce. For this reason, the algorithms and key sizes used in the Web PKI need to migrate over time. A reasonable choice of algorithm or key size needs to be reevaluated periodically, and a transition may be needed before the expected lifetime expires.
The browser vendors have been trying to manage algorithm and key size transitions, but a long-lived trust anchor or intermediate CA certificate can have a huge number of subordinate certificates. So, removing one because it uses a weak cryptographic algorithm or a short public key can have a significant impact.
As a result, some valid trust anchors and certificates contain cryptographic algorithms long after weaknesses have been discovered and widely known. Similarly, valid trust anchors and certificates contain public keys after computational resources available to attackers have rendered them too weak. We have seen a very successful migration away from certificates that use the MD2 or MD5 one-way hash functions. However, there are still a great number of certificates that use SHA-1 and 1024-bit RSA public keys, and these should be replaced.
Today, the algorithms and key sizes used by a CA to sign certificates with a traditional lifespan should offer 112 to 128 bits of security. SHA-256 is a widely studied one-way hash function that meets this requirement. RSA with a public key of at least 2048 bits or ECDSA with a public key of at least 256 bits are widely studied digital signature algorithms that meet this requirement.
Obviously, additional algorithm transitions will be needed in the future as these algorithms age. These algorithms, like the ones that were used earlier, will become weaker with time. RFC 7696 [RFC7696] offers some guidelines regarding cryptographic algorithm agility.
Several years ago, many browsers did not perform certificate status checks by default. That is, browsers did not check whether the issuing CA had revoked the certificate unless the user explicitly adjusted a setting to enable this feature. This check can be made by fetching the most recent certificate revocation list (CRL) RFC 5280 [RFC5280], or this check can use the Online Certificate Status Protocol (OCSP) RFC 6960 [RFC6960]. The location of the CRL or the OCSP responder is usually found in the certificate itself. However, both of these approaches add latency. The desire to provide a responsive user experience is a significant reason that this feature has not been turned on by default. Mobile browsers simply do not bother to check revocation status [IMC2015].
Certificate status checking needs to be used at all times. Several techniques have been tried by CAs and browsers to make certificate status checking more efficient. Many CAs are using Content Delivery Networks (CDNs) to deliver CRLs and OCSP responses, resulting in very high availability and low latency. Yet, browser vendors are still reluctant to perform standard-based status checking by default for every session.
Measurements in 2015 [IMC2015] show that a surprisingly large fraction of Web PKI certicates have been revoked. These same measurements show that browsers are not obtaining current certificate revocation information because it is too expensive in terms of latency and bandwidth. Finally, only a small number of CRL and OCSP servers are available over IPv6, and as more of the Web moves to IPv6 [ABLOG] this is expected to become an increasingly significant issue.
The following subsections identify some approaches for reducing the perceived and actual cost of revocation status checks.
Short-lived certificates are an excellent way to reduce the need for certificate status checking. The shorter the life of the certificate, the less time there is for anything to go wrong. If the lifetime is short enough, policy might allow certificate status checking can be skipped altogether. In practice, implementation of short-lived certificates requires automation to assist web server administrators, which is a topic that is discussed elsewhere in this document.
The certificate revocation list distribution point (CRLDP) certificate extension RFC 5280 [RFC5280] allows a CA to control the maximum size of the CRLs that they issue. This helps in two ways. First, the amount of storage needed by the browser to cache CRLs is reduced. Second, and more importantly, the amount of time it takes to download a CRL can be scoped, so that the amount of time needed to fetch any single CRL is reasonable.
Few CAs take advantage of the CRLDP certificate extension to limit the size of CRLs. In fact, there are several CAs that publish extremely large CRLs. Browsers never want to suffer the latency associated with large CRLs, and some ignore the CRLDP extension when it is present. Browsers tend to avoid the use of CRLs altogether.
Some browser vendors provide a proprietary mechanism for revocation checking. These mechanisms obtain revocation status information once per day for the entire Web PKI in a very compact form. No network traffic is generated at the time that a certificate is being validated, so there is no latency associated with revocation status checking. The browser vendor infrastructure performs daily checks of the Web PKI, and then the results are assembled in a proprietary format and made available to the browser. These checks only cover the trust anchor store for that browser vendor, so any trust anchors added by the user cannot be checked in this manner.
Measurements in 2015 [IMC2015] show that proprietary status checking is not currently providing adequate coverage of the Web PKI.
Browsers can avoid transmission of CRLs altogether by using the Online Certificate Status Protocol (OCSP) RFC 6960 [RFC6960] to check the validity of web server certificates. The TLS Certificate Status Request extension is defined in Section 8 of RFC 6066 [RFC6066]. In addition, RFC 6961 [RFC6961] defines the TLS Multiple Certificate Status Request extension, which allows the web server to provide status information about its own certificate and also the status of intermediate certificates in the certification path. By including this extension in the TLS handshake, the browser asks the web server to provide an OCSP response in addition to its certificate. This approach greatly reduces the number of round trips by the browser to check the status of each certificate in the path. In addition, the web server can cache the OCSP response for a period of time, avoiding additional latency. Even in the cases where the web server needs to contact the OCSP responder, the web server usually has a higher bandwidth connection than the browser to do so.
The provision of the time-stamped OCSP response in the TLS handshake is referred to as "stapling" the OCSP response to the TLS handshake. If the browser does not receive a stapled OCSP response, it can contact the OCSP responder directly. In addition, the MUST_STAPLE feature [TLSFEATURE] can be used to insist that OCSP stapling be used.
When every browser that connects to a high volume website performs its own OCSP lookup, the OCSP responder must handle a real-time response to every browser. OCSP stapling can avoid enormous volumes of OCSP requests for certificates of popular websites, so stapling can significantly reduce the cost of providing an OCSP service.
OCSP stapling can also improve user privacy, since the web server, not the browser, contacts the OCSP responder. In this way, the OCSP responder is not able to determine which browsers are checking the validity of certificate for websites.
Many web site are taking advantage of OCSP sampling. At the time of this writing, browser venders report that about 12% the the transactions use OCSP stapling, and the number is on the rise.
All of the CAs in the trust store are equally trusted for the entire domain name space, so any CA can issue for any domain name. In fact, there have been certificates issued by CAs that are surprising to the legitimate owner of a domain. The domain name owner is surprised because they did not request the certificates. They are initially unaware that a CA has issued a certificate that contains their domain name, and once the surprising certificate is discovered, it can be very difficult for the legitimate domain name owner to get it revoked. Further, browsers and other relying parties cannot distinguish a certificate that the legitimate domain name owner requested from a surprising one.
Since all of the CAs in the trust store are equally trusted, any CA can issue a certificate for any domain name. There are known cases where attackers have thwarted the CA protections and issued certificates that were then used to mount attacks against the users of that web site [FOXIT]. For this reason, all of the CAs listed in the trust store must be very well protected.
The Baseline Requirements produced by the CA/Browser Forum [CAB2014] tell CAs the checks that need to be performed before a certificate is issued. In addition, WebTrust [WEBTRUST] provides the audit requirements for CAs, and browser vendors will remove a CA from the trust anchor store if successful audit reports are not made available.
When a CA issues a certificate to a subordinate CA, the inclusion of widely supported certificate extensions can reduce the set of privileges given to the sub-CA. This aligns with the traditional security practice of least privilege, where only the privileges needed to perform the envisioned tasks are provided. However, many sub-CAs have certificates that pass along the full powers of the CA, creating additional high-pay-off targets for attackers, and these sub-CAs may not be held to the same certificate issuance requirements and audit requirement as the parent CA.
Some major implementations have not fully implemented the mechanisms necessary to reduce sub-CA privileges. For example, RFC 5280 [RFC5280] includes the specification of name constraints, and the CA/Browser Forum guidelines [CAB2014] encourage the use of dNSNames in permittedSubtrees within the name Constraints extension. Despite this situation, one major browser does not support name constraints, and as a result, CAs are reluctant to use them. Further, global CAs are prepared to issue certificates within every top-level domain, including ones that are newly-approved. It is not practical for these global CAs to use name constraints in their sub-CA certificates.
As a result of procedural failures or attacks, surprising certificates are being issued. Several mechanisms have been defined to avoid the issuance of surprising certificates or prevent browsers from accepting them.
The Certificate Authority Authorization (CAA) [RFC6844] DNS resource record allows a domain administrator to specify one or more CAs authorized to issue certificates that include that domain name. Then, a trustworthy CA will refuse to issue a certificate for a domain name that has a CAA resource record that does not explicitly name the CA.
To date, only one major CA performs this check, and there is no indication that other CAs are planning to add this check in the near future.
HTTP Public Key Pinning (HPKP) [RFC7469] allows a web server to instruct browsers to remember the server's public key fingerprints for a period of time. The fingerprint is a one-way hash of the subject public key information in the certificate. The Public-Key-Pins header provides a maximum retention period, fingerprints of the web server certificate, and optionally fingerprints for backup certificates. The act of saving the fingerprints is referred to as "pinning". During the pin lifetime, browsers require that the same web server present a certificate chain that includes a public key that matches one of the retained fingerprints. This prevents impersonation of the website with a surprising certificate.
A website can choose to pin the CA certificate so that the browser will accept only valid certificates for the website domain that are issued by that CA. Alternatively, the website can choose to pin their own certificate and at least one backup certificate in case the current certificate needs to be replaced due to a compromised server.
Some browser vendors also pin certificates by hardcoding fingerprints of very well known websites.
When HPKP is used, browsers may be able to detect a man-in-the-middle. Sometimes the man-in-the-middle is an attacker, and other times a service provider purposefully terminates the TLS at a location other than the web server. One example became very public in February 2012 when Trustwave admitted that it had issued a subordinate CA certificate for use by a company to inspect corporate network traffic [LC2012]. When HPKP is used, the browser user will be notified if the key-pining is violated, unless the violating certificate can be validated to a locally installed trust anchor. In this situation, the browser is assuming that the user intended to explicitly trust the certificate.
HTTP Strict Transport Security (HSTS) [RFC6797] is a security policy mechanism that protects secure websites against downgrade attacks, and it greatly simplifies protection against cookie hijacking. The presence of the Strict-Transport-Security header tells browsers that all interactions with this web server should never use HTTP without TLS, providing protection against eavesdropping and active network attacks.
When a web server includes the Strict-Transport-Security header, the browser is expected to do two things. First, the browser automatically turns any insecure links into secure ones. For instance, "http://mysite.example.com/mypage/" will be changed to "https://mysite.example.com/mypage/". Second, if the TLS Handshake results in some failure, such as the certificate cannot be validated, then an error message is displayed and the user is denied access to the web application.
The DNS-Based Authentication of Named Entities (DANE) [RFC6698] allows domain administrators to specify the raw public keys or certificates that are used by web servers in their domain. DANE leverages the DNS Security Extensions (DNSSEC) [RFC4034][RFC4035], which provides digital signatures over DNS zones that are validated with keys that are bound to the domain name of the signed zone. The keys associated with a domain name can only be signed by a key associated with the parent of that domain name. For example, the DNSSEC keys for "www.example.com" can only be signed by the DNSSEC keys for "example.com". Therefore, a malicious actor can only compromise the keys of their own subdomains. Like the Web PKI, DNSSEC relies on public keys used to validate chains of signatures, but DNSSEC has a single root domain as opposed to a multiplicity of trusted CAs.
DANE binds raw public keys or certificates to DNS names. The domain administrator is the one that vouches for the binding of the public key or the certificate to the domain name by adding the TSLA records to the zone and then signing the zone. In this way, the same administrator is responsible for managing the DNS names themselves and associated public keys or certificates with those names. DANE restricts the scope of assertions that can be made, forcing them to be consistent with the DNS naming hierarchy.
In addition, DNSSEC reduces opportunities for redirection attacks by binding the domain name to the public key or certificate.
Some Web PKI certificates are being posted in TLSA records, but browsers expect to receive the server certificate in the TLS handshake, and there is little incentive to confirm that the received certificate matches the one posted in the DNS. For this reason, work has begun on a TLS extension that will allow the DNSSEC-protected information to be provided in the handshake, which will eliminate the added latency [TLSCHAIN].
Certificate Transparency (CT) [RFC6962] offers a mechanism to detect surprising certificates, and once detected, administrators and CAs can take the necessary actions to revoke the surprising certificates.
When requesting a certificate, the administrator can request the CA to include an embedded Signed Certificate Timestamp (SCT) in the certificate to ensure that their legitimate certificate is logged with one or more CT logs.
An administrator, or another party acting on behalf of the administrator, is able to monitor one or more CT logs to which a pre-certificate or certificate is submitted, and detect the logging of a pre-certificate or certificate that contains their domain name. When such a pre-certificate or certificate is detected, the CA can be contacted to to get the surprising certificate revoked.
In the future, a browser may choose to reject certificates that do not contain an SCT, and potentially notify the website administrator or CA when they encounter such a certificate. Such reporting will help detect mis-issuance of certificates and lead to their revocation.
There have been several attempts to provide automation for routine tasks that are performed by web server administrators, such as certificate renewal. For example, some commercial tools offer automated certificate renewal and installation [DCEI][SSLM]. Also, at least one proposal was brought to the IETF that allows a web server to automate obtaining and renewing certificates [PHBOB]. Without automation, there are many manual steps involved in getting a certificate from a CA, and to date none of these attempts at automation have enjoyed widespread interoperability and adoption. There are at least two ways that this impacts web security. First, many web sites do not have a certificate at all. The cost, time, and effort are too great for the system administrator. This is especially true if the web site is not involved in financial transactions or some other critical activity. Second, once a certificate is obtained, a replacement is not obtained until the current one expires. Automation can reduce the amount of time that an administrator needs to dedicate to certificate management, and it can make certificate renewal timely and automatic. Both of these should lead to more widespread deployment and improved web security.
The IETF ACME working group [ACMEWG] is working on protocols that will provide system administrators with an automated way to enroll and renew their certificates. The expectation is that these specifications will lead to widely available and interoperable tools for system administrators. The expectation is that these protocols and tools will be supported by all web server environments and CAs, which will greatly reduce complexity and cost. In addition, the easier renewal process provided by automation can be used to reduce certificate lifetimes, which in turn will reduce the time required to flush old algorithms out of the system when it is decided to transition to newer more secure algorithms.
As with many technologies, the issues and complexities associated with Web PKI use and deployment are just as much policy and process as technical. These have evolved over time as well. This section discusses the ways that business models and operational policies and processes impact the Web PKI.
A very basic question for users of the Web PKI is "Who do you trust?" The system for determining which CAs are added to or removed from the trust store in browsers has been perceived by some as opaque and confusing. As mentioned earlier, the CA/Browser Forum has developed baseline requirements for the management and issuance of certificates [CAB2014] for individual CAs. However, the process by which an individual CA gets added to the trust store for each of the major browsers is not straightforward. The individual browser vendors determine what should and should not be trusted by including those trusted CAs in their trust store. They do this by leveraging the AICPA/CICA WebTrust Program for Certification Authorities [WEBTRUST]. This program provides auditing requirements and a trust mark for CAs. Failure to pass an audit can result in the CA being removed from the trust store.
Once the browser has shipped, how does a user know which CAs are trusted or what has changed recently? For an informed user, information about which CAs have been added to or deleted from the browser trust store can be found in the release notes. Users can also examine the policies of the various CAs which would have been developed and posted for the WebTrust Program. However, this may be considered a fairly high barrier for the average user. There are also options to make local modifications by educated users, but there is little understanding about the implications of these choices. How does an individual, organization, or enterprise really determine if a particular CA is trustworthy? Do the default choices inherited from the browser vendors truly represent the organization's trust model? What constitutes sufficiently bad behavior by a CA to cause removal from the trust store?
One form of bad behavior by CAs is the mis-issuance of certificates. This mis-issuance can be either an honest mistake by the CA, malicious behavior by the CA, or a case where an external party has duped the CA into the mis-issuance. When a CA has delegated authority to a sub-CA, and then the sub-CA issued bad certificates either unintentionally or maliciously, the CA is able to deny responsibility for the actions of the sub-CA. However, the CA may be the only party that can revoke the sub-CA certificate to protect the overall Web PKI.
Another complication with CAs and the trust store maintained by the browser vendor is an enterprise managed PKI. For example, the US Department of Defense operates its own PKI. In this case, the enterprise maintains its own PKI for the exclusive use of the enterprise itself. A bridge CA may be used to connect related enterprises. The complication in this approach is that the revocation mechanisms don't work with any additions that have been made by the enterprise. See Section 3.2.3 on proprietary revocation checks.
The guidelines provided by the WebTrust program [WEBTRUST] provide a framework for removing a CA from the trust store. However, the implications of removing a CA can be significant. There may be a few very large CAs that are critical to significant portions of Internet infrastructure. Removing one of these trusted CAs can have a significant impact on a large cross section of Internet users resulting in potentially large numbers of websites no longer being trusted. Users are already struggling to understand the implications of untrusted websites and often ignore the current warnings as discussed below.
There are a number of organizations that play significant roles in the operation of the Web PKI, including the CAB Forum, the WebTrust Program, and the browser vendors. These organizations act on behalf of the entire Internet community. Transparency in these operations is vital to basic trust in the Web PKI. As one example, in the past the CAB Forum was perceived as being a closed forum; however, some changes were made to the operational procedures to allow more visibility if not actual participation in the process [CAB1.2]. How do we ensure that these processes continue to evolve in an open, inclusive, and transparent manner? Currently, as the name implies, the CAB Forum members represent CAs and browser vendors. How do we ensure that relying parties have a voice in this forum?
Since the Web PKI is widespread, applications beyond the World Wide Web are making use of the Web PKI. For example, the Web PKI is used to secure the connections between SMTP servers. In these environments, the browser-centric capabilities are unavailable. For example, see Section 3.2.3 on proprietary revocation checks. The current governance structure does not provide a way for these other applications to participate. How do we ensure that these other applications get a voice in this forum?
Beyond the technical mechanisms that constitute the Web PKI itself, there are additional technologies that impact the success of the Web PKI infrastructure. Examples of these are discussed in this section.
Many people find browser error messages related to certificates confusing. Good man-machine interfaces are always difficult, but in this situation users are unable to understand the risks that they are accepting by clicking "okay". Users have been basically trained to ignore the warnings provided by the infrastructure rendering this warning ineffective. This aspect of browser usability needs to be improved for users to make better security choices. (Editor note: Additional detail with references is needed for this section.)
Time synchronization is another factor that impacts the security and reliability of the Web PKI. Reasonably accurate time is needed to check certificate expiration and to determine whether cached revocation status information is fresh. There is ongoing work to improve the security of the time synchronization infrastructure, and it will use certificates to authenticate time servers. Since the certificate infrastructure relies on quality time synchronization, this dependency creates a boot strapping issue.
To make the Web PKI more secure and more robust, the following priorities have been identified and are recommended for further development and deployment:
In addition, solutions to these procedural and policy challenges are needed:
This document considers the weaknesses of the current Web PKI system and provides recommendations for improvements. Some of the risks associated with doing nothing or continuing down the current path are articulated. The Web PKI is a vital component of a trusted Internet and as such needs to be improved to sustain continued growth of the Internet.
None.
{{{ RFC Editor: Please remove this section prior to publication. }}}
| [RFC5280] | Cooper, D., Santesson, S., Farrell, S., Boeyen, S., Housley, R. and W. Polk, "Internet X.509 Public Key Infrastructure Certificate and Certificate Revocation List (CRL) Profile", RFC 5280, DOI 10.17487/RFC5280, May 2008. |
| [RFC6960] | Santesson, S., Myers, M., Ankney, R., Malpani, A., Galperin, S. and C. Adams, "X.509 Internet Public Key Infrastructure Online Certificate Status Protocol - OCSP", RFC 6960, DOI 10.17487/RFC6960, June 2013. |
This document has been developed within the IAB Privacy and Security Program. The authors greatly appreciate the review and suggestions provided by Rick Andrews, Mary Barnes, Richard Barnes, Marc Blanchet, Alissa Cooper, Nick Doty, Stephen Farrell, Joe Hall, Ted Hardie, Ralph Holz, Lee Howard, Christian Huitema, Eliot Lear, Xing Li, Lucy Lynch, Gervase Markham, Andrei Robachevsky, Thomas Roessler, Jeremy Rowley, Christine Runnegar, Jakob Schlyter, Wendy Seltzer, Brian Trammell, and Juan Carlos Zuniga.
{{{ RFC Editor: Please add the names to the IAB members at the time that this document is put into the RFC Editor queue. }}}