| Internet-Draft | Best practices for password hashing and | April 2021 |
| Whited | Expires 8 October 2021 | [Page] |
This document outlines best practices for handling user passwords and other authenticator secrets in client-server systems making use of SASL.¶
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 https://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 8 October 2021.¶
Copyright (c) 2021 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 (https://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.¶
Following best practices when hashing and storing passwords for use with SASL impacts a great deal more than just a user's identity. It also affects usability, backwards compatibility, and interoperability by determining what authentication and authorization mechanisms can be used.¶
Various security-related terms are to be understood in the sense defined in [RFC4949]. Some may also be defined in [NISTSP63-3] Appendix A.1 and in [NISTSP132] section 3.1.¶
Throughout this document the term "password" is used to mean any password, passphrase, PIN, or other memorized secret.¶
Other common terms used throughout this document include:¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
For clients and servers that support password based authentication using SASL [RFC4422] it is RECOMMENDED that the following mechanisms be implemented:¶
System entities SHOULD NOT invent their own mechanisms that have not been standardized by the IETF or another reputable standards body. Similarly, entities MUST NOT implement any mechanism with a usage status of "OBSOLETE", or "LIMITED", or "MUST NOT be used" in the IANA SASL Mechanisms Registry [IANA.sasl.mechanisms]. For example, entities MUST NOT implement DIGEST-MD5 (deprecated in [RFC6331]).¶
The SASL mechanisms discussed in this document do not negotiate a security layer. Because of this a strong security layer such as TLS [RFC8446] MUST be negotiated before SASL mechanisms can be advertised or negotiated.¶
Clients often maintain a list of preferred SASL mechanisms, generally ordered by perceived strength to enable strong authentication. To prevent downgrade attacks by a malicious actor that has successfully executed an in-the-middle attack on a connection, or compromised a trusted server's configuration, clients SHOULD implement "mechanism pinning". That is, after the first successful authentication with a strong mechanism, clients SHOULD make a record of the authentication and thereafter only advertise and use mechanisms of equal or higher perceived strength.¶
The following mechanisms are ordered by their perceived strength from strongest to weakest with mechanisms of equal strength on the same line. The remainder of this section is merely informative. In particular this example does not imply that mechanisms in this list should or should not be implemented.¶
The EXTERNAL mechanism defined in [RFC4422] appendix A is placed at the top of the list. However, its perceived strength depends on the underlying authentication protocol. In this example, we assume that TLS [RFC8446] services are being used.¶
The channel binding ("-PLUS") variants of SCRAM [RFC5802] are listed above their non-channel binding cousins, but may not always be available depending on the type of channel binding data available to the SASL negotiator.¶
Finally, the PLAIN mechanism sends the username and password in plain text and therefore requires a strong security layer such as TLS for the password to be protected in transit. However, if the server is trusted to know the password PLAIN does allow for the use of a strong key derivation function (KDF) for storing the authentication data at rest and provides for password hash agility.¶
Clients SHOULD always store authentication secrets in a trusted and encrypted keystore such as the system keystore, or an encrypted store created specifically for the clients use. They SHOULD NOT store authentication secrets as plain text.¶
If clients know that they will only ever authenticate using a mechanism such as SCRAM [RFC5802] where the original password is not needed after the first authentication attempt they SHOULD store the SCRAM bits or the hashed and salted password instead of the original password. However, if backwards compatibility with servers that only support the PLAIN mechanism or other mechanisms that require using the original password is required, clients MAY choose to store the original password so long as an appropriate keystore is used.¶
Servers MUST NOT support any mechanism that would require authentication secrets to be stored in such a way that they could be recovered in plain text from the stored information. This includes mechanisms that store authentication secrets using reversable encryption, obsolete hashing mechanisms such as MD5 or hashing mechanisms that are cryptographically secure but designed for speed such as SHA256.¶
Servers MUST always store passwords only after they have been salted, peppered (if possible with the given authentication mechanism), and hashed using a strong KDF. A distinct salt SHOULD be used for each user, and each SCRAM family supported. Salts SHOULD be generated using a cryptographically secure random number generator. The salt MAY be stored in the same datastore as the password. A pepper stored in the application configuration, or a secure location other than the datastore containing the salts, SHOULD be combined with the password before hashing if possible with the given authentication mechanism. Peppers SHOULD NOT be combined with the salt because the salt is not secret and may appear in the final hash output.¶
The following restrictions MUST be observed when generating salts and peppers:¶
| Parameter | Value |
|---|---|
| Minimum Salt Length | 4 bytes |
| Minimum Pepper Length | 14 bytes |
When authenticating using PLAIN or similar mechanisms that involve transmitting the original password to the server the password MUST be hashed and compared against the salted and hashed password in the database using a constant time comparison.¶
Each time a password is changed a new random salt MUST be created and the iteration count and pepper (if applicable) MUST be updated to the latest value required by server policy.¶
If a pepper is used, consideration should be taken to ensure that it can be easily rotated. For example, multiple peppers could be stored. New passwords and reset passwords would use the newest pepper and a hash of the pepper using the same KDF that was used on the password could then be stored in the database next to the salt so that future logins can identify which pepper in the list was used. This is just one example, pepper rotation schemes are outside the scope of this document.¶
When properly configured, the following commonly used KDFs create suitable password hash results for server side storage. The recommendations in this section may change depending on the hardware being used and the security level required for the application.¶
With all KDFs proper tuning is required to ensure that it meets the needs of the specific application or service. For persistent login an iteration count or work factor that adds approximately a quarter of a second to login may be an acceptable tradeoff since logins are relatively rare. By contrast, verification tokens that are generated many times per second may need to use a much lower work factor.¶
Argon2 [ARGON2ESP] is the 2015 winner of the Password Hashing Competition and the current OWASP recommendation for password hashing. Security considerations, test vectors, and parameters for tuning argon2 can be found in [I-D.irtf-cfrg-argon2]. The defaults are copied here for easier reference.¶
| Parameter | Value |
|---|---|
| Degree of parallelism (p) | 4 |
| Minimum memory size (m) | 2 GiB |
| Minimum number of iterations (t) | 1 |
| Algorithm type (y) | Argon2id (2) |
| Minimum output length | 32 |
bcrypt [BCRYPT] is a Blowfish-based KDF.¶
| Parameter | Value |
|---|---|
| Recommended Cost | 12 |
| Maximum Password Length | 50-72 bytes depending on the implementation |
PBKDF2 [RFC8018] is used by the SCRAM [RFC5802] family of SASL mechanisms.¶
| Parameter | Value |
|---|---|
| Minimum iteration count (c) | 310,000 |
| Hash | HMAC-SHA256 |
| Output length (dkLen) | min(hLen, 32) (where hLen is the length of the chosen hash) |
When PBKDF2 is used with HMAC such as in the SCRAM [RFC5802] family of SASL mechanisms the password is pre-hashed if it is longer than the block size of the hash function (hLen, or 64 bytes for SHA-256). Care should be taken to ensure that the implementation of PBKDF2 does this before the iterations, otherwise long hashes may become significantly more expensive than expected, possibly resulting in a Denial-of-Service (DOS).¶
The [SCRYPT] KDF is designed to be memory-hard and sequential memory-hard to prevent against custom hardware based attacks.¶
Security considerations, test vectors, and further notes on tuning scrypt may be found in [RFC7914].¶
| Parameter | Value |
|---|---|
| Minimum CPU/Memory cost parameter (N) | 32768 (N=2^15) |
| Blocksize (r) | 8 |
| Parallelization parameter (p) | 1 |
| Minimum output length (dkLen) | 32 |
Before any other password complexity requirements are checked, the preparation and enforcement steps of the OpaqueString profile of [RFC8265] SHOULD be applied (for more information see the Internationalization Considerations section). Entities SHOULD enforce a minimum length of 8 characters for user passwords. If using a mechanism such as PLAIN where the server performs hashing on the original password, a maximum length between 64 and 128 characters MAY be imposed to prevent denial of service (DoS) attacks. Entities SHOULD NOT apply any other password restrictions.¶
In addition to these password complexity requirements, servers SHOULD maintain a password blocklist and reject attempts by a claimant to use passwords on the blocklist during registration or password reset. The contents of this blocklist are a matter of server policy. Some common recommendations include lists of common passwords that are not otherwise prevented by length requirements, and passwords present in known breaches.¶
The PRECIS framework (Preparation, Enforcement, and Comparison of Internationalized Strings) defined in [RFC8264] is used to enforce internationalization rules on strings and to prevent common application security issues arrising from allowing the full range of Unicode codepoints in usernames, passwords, and other identifiers. The OpaqueString profile of [RFC8265] is used in this document to ensure that codepoints in passwords are treated carefully and consistently. This ensures that users typing certain characters on different keyboards that may provide different versions of the same character will still be able to log in. For example, some keyboards may output the full-width version of a character while other keyboards output the half-width version of the same character. The Width Mapping rule of the OpaqueString profile addresses this and ensures that comparison succeeds and the claimant is able to be authenticated.¶
When enforcing a minimum password length the authentication server SHOULD NOT count bytes as single Unicode scalar values may take up many bytes. Similarly, a single emoji may be constructed from many Unicode scalar values, so it may not be appropriate to count scalar values or code points. Instead consider counting the number Grapheme Clusters as defined in [UAX29].¶
This document contains recommendations that are likely to change over time. It should be reviewed regularly to ensure that it remains accurate and up to date. Many of the recommendations in this document were taken from [OWASP.CS.passwords], [NISTSP63b], and [NISTSP132].¶
The "-PLUS" variants of SCRAM [RFC5802] support channel binding to their underlying security layer, but lack a mechanism for negotiating what type of channel binding to use. In [RFC5802] the tls-unique [RFC5929] channel binding mechanism is specified as the default, and it is therefore likely to be used in most applications that support channel binding. However, in the absence of the TLS extended master secret fix [RFC7627] and the renegotiation indication TLS extension [RFC5746] the tls-unique and tls-server-endpoint channel binding data can be forged by an attacker that can MITM the connection. Before advertising a channel binding SASL mechanism, entities MUST ensure that both the TLS extended master secret fix and the renegotiation indication extension are in place and that the connection has not been renegotiated.¶
For TLS 1.3 [RFC8446] no channel binding types are currently defined. Channel binding SASL mechanisms MUST NOT be advertised or negotiated over a TLS 1.3 channel until such types are defined.¶
This document has no actions for IANA.¶
The author would like to thank the civil servants at the National Institute of Standards and Technology for their work on the Special Publications series. U.S. executive agencies are an undervalued national treasure, and they deserve our thanks.¶
Thanks also to Cameron Paul, Thomas Copeland, Robbie Harwood, Jim Fenton, and Alexey Melnikov for their reviews and suggestions.¶