Benjamin Kaduk has entered the following ballot position for draft-ietf-regext-secure-authinfo-transfer-06: Discuss
When responding, please keep the subject line intact and reply to all email addresses included in the To and CC lines. (Feel free to cut this introductory paragraph, however.) Please refer to https://www.ietf.org/iesg/statement/discuss-criteria.html for more information about DISCUSS and COMMENT positions. The document, along with other ballot positions, can be found here: https://datatracker.ietf.org/doc/draft-ietf-regext-secure-authinfo-transfer/ ---------------------------------------------------------------------- DISCUSS: ---------------------------------------------------------------------- (1) RFC 4086 does not state that "a high-security password must have at least 49 bits of randomness or entropy" as is claimed in Section 4.1 of this document. It merely says that so much entropy is needed to have a one-in-a-billion chance of success for successfully guessing in the model laid out, and makes no statement about (absolute) "high" security. I don't think we need to spend as much time on what RFC 4086 says as we currently do, and could probably get to the "use at least 128 bits of entropy" advice much sooner. (2) There's also some text in Section 5.3 that I'd like to discuss briefly: The registry MUST NOT return any indication of whether the authorization information is set or unset to the non-sponsoring registrar by not returning the authorization information element in the response. The registry MAY return an indication to the sponsoring registrar that the authorization information is set by using an empty authorization information value. The registry MAY return an indication to the sponsoring registrar that the authorization information is unset by not returning the authorization information element. This seems to be assigning semantics to both absent-authinfo and empty-authinfo in the <info> response, but is giving *different* semantics to the response-to-sponsoring-registrar and response-to-non-sponsoring-registrar cases. Is there precedent for changing the semantics of the response based on the identity of the client like this (not just changing the content of the response)? Can we come up with a scheme that provides consistent semantics to all clients, perhaps based on <domain:null> vs empty <domain:pw> for unset/set, leaving "element is absent" for the deliberately ambiguous case? (3) We may also need to discuss the efficacy of the transition plan, per my comments in Sections 6.1 and 6.3 -- my current understanding is that the proposed plan will break some existing workflows. I am not sure if that is intended, desirable, and/or tolerable, and welcome further insight. ---------------------------------------------------------------------- COMMENT: ---------------------------------------------------------------------- I think that the introduction would benefit from expanding on the high-level motivation for this work (which I assume to include): the current/original lifecycle for authorization information involves long-term storage of encrypted (not hashed) passwords, which presents a significant latent risk of password compromise and is not consistent with current best practices. The mechanisms in this document provide a way to avoid long-term password storage entirely, and to only require the storage of hashed (not retrievable) passwords instead of encrypted passwords. (Or, in a more colloquial language, "passwords suck, and we want to get out of the business of handling them to the extent possible".) The third paragraph does talk about the "overall goal", but doesn't say much about what we're moving *away* from (and why). I think giving some explicit consideration of the lifecycle and protocol interactions for the password 'salt' would be helpful. (That is, that it's picked at random by the registry per password when a password is set and never goes on the wire, but is stored alongside the hashed password.) The treatment in the introduction of "[use] the existing features of the EPP RFCs" made me wonder why this needed to be on the standards-track, as opposed to being an informational description of how to use what's already there. The actual core of the spec, which includes changes to the semantics of some XML elements (e.g., in the info response), is clearly protocol work, though, so perhaps the abstract/introduction could be revisited to clarify the scope of the work. Section 1 "Strong Random Authorization Information": The EPP RFCs define the password-based authorization information value using an XML schema "normalizedString" type, so they don't restrict what can be used in any way. This operational practice defines the I suggest s/in any way/in any substantial way/ (not being able to use CR/LF/TAB is in some sense a restriction). "Short-Lived Authorization Information": The EPP RFCs don't [...] upon a successful transfer. All of these features can be supported by the EPP RFCs. They can be supported, sure, but what about in practice? Can we rely on such functionality being present? Section 3 namespace URI in the login and greeting extension services. The namespace URI "urn:ietf:params:xml:ns:epp:secure-authinfo-transfer- 1.0" is used to signal support for the operational practice. The Written in this way this is codepoint squatting, assuming that the requested XML namespace value will be assigned. Given that Section 8 implies this stuff is deployed already, there really should have been an early allocation made. A client that receives the namespace URI in the server's Greeting extension services, can expect the following supported behavior by the server: 1. Support an empty authorization information value with a create command. [...] It's interesting to compare this to RFC 5731, that says "Authorization information as described in Section 2.6 is REQUIRED to create a domain object. [...] Failure to protect authorization information from inadvertent disclosure can result in unauthorized transfer operations and unauthorized information release. [...]" In some sense we are introducing a rather significant philosophical change in the nature of authorization information, that might be called out more prominently. 7. Support automatically unsetting the authorization information upon a successful completion of transfer. Just "support", or actually will in practice? (Probably applies to some of the other enumerated points as well, for both client and server, though I think not to all of them.) Section 4 with the <contact:pw> element). Other EPP objects that support password-based authorization information for transfer can use the Secure Authorization Information defined in this document. For the This is phrased just as "can use" (not "will use"), but we are supposedly defining an XML namespace used for capability negotiation in the initial EPP exchange, which really ought to have well-specified semantics. *Must* the secure authorization information defined in this document be used for any applicable transfer, when the XML namespace we define is in effect for the capabilities to use? Section 4.1 The strength of the random authorization information is dependent on the actual entropy of the underlying random number generator. For the random number generator, the practices defined in [RFC4086] and section 4.7.1 of the NIST Federal Information Processing Standards (FIPS) Publication 140-2 [FIPS-140-2] SHOULD be followed to produce random values that will be resistant to attack. A random number generator (RNG) is preferable over the use of a pseudorandom number generator (PRNG) to reduce the predictability of the authorization information. The more predictable the random number generator is, the lower the true entropy, and the longer the required length for the authorization information. This is not really the advice that I would be giving, myself. For one, using a true RNG is not necessarily better than a PRNG, since a good PRNG will have a lot of well-thought-out "whitening" functionality that makes the output fairly uniform. A true RNG can be a true RNG while still sampling from a non-uniform distribution and leaving patterns in the output. But more importantly, implementors of EPP are highly unlikely to need to care about the entropy gathering practices specified by NIST SP 140-2 and RFC 4086 -- they can and should just use /dev/urandom! RFC 4086 was written in a time when /dev/urandom was not as reliable as it is nowadays, but the advice of "to obtain randon numbers under Linux, Solaris, [...] all an application has to do is open either /dev/random or /dev/urandom and read the desired number of bytes" is arguably the most important guidance in it. (There's also a corresponding Windows API.) If we don't think people will want to implement a 94-character alphabet themselves, we can suggest the widely available base64 encoding and say something like "read 18 bytes from /dev/random and base64 encode it, which will produce 24 characters of encoded output" or give the ROUNDUP(128/log2 64) math. (I use 18 bytes because that avoids base64 padding characters.) Section 4.2 I strongly suggest giving some guidance to registrars on how to set the TTL (presumably a week or a few days is doable for common/generic domain transfers?). The requirements in ยง4.3 are not really aligned with current best practices for password hashing for long term storage (which, admittedly, are designed for human-selected passwords and not random ones), so clamping down the TTL is going to be helpful for putting bounds on some classes of attack. Section 4.3 I note that draft-ietf-kitten-password-storage is underway to write down best practices for password hashing and storage. It is probably not mature enough to be used as a definitive reference yet, but could be useful information. 1. The authorization information MUST be stored by the registry using a strong one-way cryptographic hash, with at least a 256-bit hash function such as SHA-256 [FIPS-180-4], and with a random salt. Typical password-hashing recommendations these days are things like Argon2 (winner of https://www.password-hashing.net/) or PBKDF2 (for somewhat more legacy systems). The iteration count (and other parameters, for Argon2) can be tweaked depending on the systems in question and the esteimated strength of the password, but nobody is doing just a single-iteration of (salted) SHA256. I note that we recently approved draft-ietf-lamps-crmf-update-algs that leaves in place a requirement that the iteration count MUST be at least 100 (and adds "SHOULD be as large as server performance will allow, typically at least 10,000"). If the intent is to take advantage of the special considerations here about the nature of the passwords in question, in order to diverge from password-hashing best practice, we should be explicit about the intentional divergence. 5. The plain text version of the authorization information MUST NOT be written to any logs by the registrar or the registry, nor nit: "the registrar" is perhaps ambiguous in this scenario where we have both losing and gaining registrars. Section 4.4 1. Any input authorization information value MUST NOT match an unset authorization information value. Does this only apply to non-empty input authorization information? 3. A non-empty input authorization information value MUST be hashed and matched against the set authorization information value, which is stored using the same hash algorithm. It might be worth a few sentences (not necessarily here) about password-hashing-algorithm agility and what an algorithm transition would look like. Section 5.2 Because of this, registries may validate the randomness of the authorization information based on the length and character set required by the registry. For example, validating an authorization value contains a combination of upper-case, lower-case, and non- alphanumeric characters, in an attempt to assess the strength of the value, and return an EPP error result of 2202 if the check fails. Such checks are, by their nature, heuristic and imperfect, and may identify well-chosen authorization information values as being not sufficiently strong. Registrars, therefore, must be prepared for an error response of 2202, "Invalid authorization information", and respond by generating a new value and trying again, possibly more than once. I note for the record that we had an earlier conversation about this behavior, and I still believe that it does not reflect a best practice for minimizing the use of weak passwords. That said, it is a non-normative example, and we basically already had our discussion on this topic, so there is no need to rehash it again -- this is a non-blocking comment. Section 6 3. Losing registrar retrieves the stored authorization information locally or queries the registry for authorization information nit: I think s/retrieves the stored authorization information locally/retrieves the locally stored authorization information/ helps readability. Section 6.1 How do these features interact with the presence (or absence) of the secure-authinfo-transfer XML namespace in the <login>/greeting exchange? It seems like at least the "don't return the authorization information in the info response" change, if unilaterally implemented by the registry, would break the classic workflow for registrars that do not store the authorization information locally and require retrieving it from the registry. (Or are they required to implement the ability to re-set the authorization information with an update, so that recovery is possible?) Section 6.2 Hash New Authorization Information Values: Change the create command and the update command to hash instead of encyrpting the nit: s/encyrpting/encrypting/ Supporting Comparing Against Encrypted and Hashed Authorization Information: Change the info command and the transfer request command to be able to compare a passed authorization information value with either a hashed or encyrpted authorization information value. This seems to leave it implicit that the stored values in the registry include an indication of whether they are encrypted or hashed. This is probably trivial to ensure, just by virtue of being formatted differently, but is an important enough property that I would suggest mentioning it specifically. Section 6.3 As for the case in Section 6.1, are these changes contingent on the negotiation of the use of secure-authinfo-transfer? Disallowing the creation of entries with non-empty authorization information values seems like it would break existing clients that do not implement secure-authinfo-transfer. Is there some mechanism in place that is going to make secure-authinfo-transfer (e.g., contractually) required to implement for registrars? Section 9 Section 4.1 defines the use a secure random value for the generation of the authorization information. The server SHOULD define policy related to the length and set of characters that are included in the randomization to target the desired entropy level, with the recommendation of at least 128 bits for entropy. The authorization information server policy is communicated to the client using an out- of-band process. The client SHOULD choose a length and set of characters that results in entropy that meets or exceeds the server policy. A random number generator (RNG) is preferable over the use of a pseudorandom number generator (PRNG) when creating the authorization information value. [my comment from above about RNG vs PRNG applies here as well.] I am a little uneasy about the "SHOULD define policy", which may just reflect a misundersanding of the text. If this is just an administrative policy that is written and read by humans, that is probably useful, but how it interacts with mechanical enforcement (per my previous comments) could spill over into a regime of giving normative recommendations to do things that I do not believe are best practice. Section 4.2 defines the use of an authorization information Time-To- Live (TTL). The registrar SHOULD only set the authorization information during the transfer process by the server support for I guess I don't understand why this is only a SHOULD (and this same requirement seems to appear in Section 4.2 as well). Given that we propose for the registry to reject creation with non-empty authorization information, it doesn't seem too big of a change to require that registrars also comply with this workflow. the end of the transfer process. The registry MUST store the authorization information using a one-way cryptographic hash of at least 256 bits and with a random salt. All communication that [this text might have to change if the earlier comment about password hashing techniques results in a textual change] Section 11.1 (If my suggestion about RNG guidance is accepted, FIPS-140-2 will no longer need to be a normative reference.) I don't really understand why RFC 3688 is listed as normative but RFC 7451 is listed as informative -- on both cases they're only referenced as the specification that created an IANA registry that we're getting an allocation in. _______________________________________________ regext mailing list regext@ietf.org https://www.ietf.org/mailman/listinfo/regext
