Re: [TLS] draft-ietf-tls-tls13-26 is vulnerable to externally set PSK identity enumeration

[Eric Rescorla <[e-mail]>]

On Mon, Mar 19, 2018 at 1:33 PM, Nikos Mavrogiannopoulos <[e-mail]>
wrote:

> On Fri, 2018-03-16 at 14:45 -0500, Benjamin Kaduk wrote:
> > On Fri, Mar 16, 2018 at 09:11:32AM -0400, Christian Huitema wrote:
> > >
> > >
> > > On 3/15/2018 5:51 PM, Benjamin Kaduk wrote:
> > > > On Thu, Mar 15, 2018 at 12:25:38PM +0100, Hubert Kario wrote:
> > > > ...
> > > > > we do not have a reliable mechanism of differentiating between
> > > > > external and
> > > > > resumption PSKs while parsing Client Hello
> > > >
> > > > Well, a valid external PSK (identity) the server will of course
> > > > recognize, and we have a SHOULD-level requirement that the
> > > > obfuscated_ticket_age is zero for external PSKs.  I haven't
> > > > gotten
> > > > to think through whether there is still potential for information
> > > > leakage about external PSK identities, but it seems like there
> > > > would
> > > > not be, provided that the server prefers resumption to external-
> > > > PSK
> > > > full handshakes.
> > > >
> > >
> > > I am concerned with the privacy issues linked to these "external
> > > PSK
> > > identities". If a system is set so that clients use static PSK
> > > identities, then the identity is an identifier and the client's
> > > movements and connections can be tracked. I don't think privacy is
> > > improved if we make it easy to differentiate external identities
> > > from
> > > resumption tickets.
> >
> > Oh, of course, the privacy considerations of the current external
> > PSK scheme are terrible!  This follows naturally from external PSKs
> > having not really been a first-class citizen while we were designing
> > things; we stuffed resumption PSKs together with external PSKs for
> > the convenience of having them use the same binder construct and
> > only needing to have one extension at the end of the ClientHello.
> > Resumption flows get single-use tickets for privacy preservation,
> > and external PSKs get infinite use and a gigantic correlation
> > channel.
>
> I agree.
>
> > > If you want to use PSK with some level of privacy, you might adopt
> > > a
> > > different setup. For example, servers could provision the clients
> > > with a
> > > set of single-use external PSK identities. But then, that looks a
> > > lot
> > > like resumption. Or, clients could generate single-use external PSK
> > > identities by encrypting their long term identity and a nonce with
> > > the
> > > public key of the server, a design which of course has its own set
> > > of
> > > issues.
> > >
> >
> > But, as you note, this is something of an open problem for how to do
> > well, and this document is already approved by the IESG.  (It's also
> > not entirely clear that the TLS WG would be the best place to design
> > this sort of scheme, though of course it could choose to do so.)
> >
> > So to me the open question is whether we consider enumeration of
> > external PSK identifiers to be something we can address reasonably
> > at this stage of the document's lifecycle at all.  (I also note that
> > the presence of a CVE number for a similar issue does not
> > necessarily mean anything -- CVE assignments can occur for
> > situations with no actual security impact and/or against the wishes
> > of the software authors.)  I don't think anyone has proposed a
> > minimal change that would close the enumeration channel, and given
> > that external PSKs already have bad privacy properties, it seems
> > like we may just have to accept this state of affairs for this
> > document.
>
> That's a good general remark, but not really the case for the
> vulnerabilities that Hubert pointed out.
>
> > Hubert also says:
> >
> > % so there's no reliable way to say that, yes, this identity is or is
> > not a
> > % resumption ticket, if I can't decrypt it
> >
> > Mostly.  External PSKs are established out of band, and that
> > provisioning process *could* include knowledge that the
> > obfuscated_ticket_age would always be zero when those PSKs are in
> > use, and that would be reliable for those specific parties.
>
> I believe that this can happen in an interoperable way if the protocol
> mandates it (which is not the case now). These specific parties may use
> software from different vendors which may use different conventions if
> the protocol is not clear enough.
>
> > It's probably also worth considering the two cases for server
> > behavior when presented with a PSK id that is neither a known
> > external PSK nor a known resumption ticket -- the server could
> > either treat it as an unknown external PSK id or as a resumption
> > ticket that fails to decrypt.  The latter case fails because the
> > attacker can try candidate external identities and the server falls
> > back to a full handshake unless the PSK ID is good.  (Well, maybe
> > the server rejects PSK IDs that are shorter than a ticket would be.)
> > The first case is not really usable since it would lead to spurious
> > triggering of the proposed "at most one external PSK" check.
> >
> > So, in addition to the "we provision external PSKs only when we know
> > that obfuscated_ticket_age will be zero", deployments could also
> > agree that external PSK ids are shorter than a given length and
> > resumption PSKs are larger, which would again provide a reliable
> > differentiator between resumption and external.
>
> That cannot easily happen. I can have multiple servers answering to the
> same hostname each using a different implementation. Any conventions
> used in one implementation would not apply to another.
>
> > % I'd really prefer we exhaust other possibilities before sacrificing
> > support
> > % for multiple external PSK. With TLS 1.2 we had ticket_hint to guide
> > PSK
> > % selection, now we're left with just server IP or hostname.
> >
> > I think that "do nothing and accept external PSK enumeration as a
> > risk" is more likely than sacrificing support for multiple external
> > PSKs, personally.
>
> The problem is that you personally are not affected by that risk and I
> guess that makes it easy for you to accept it. TLS1.2 with PSK did
> explicitly prevent enumeration (by asking implementations to proceed to
> handshake even with unknown usernames, and making up a key), meaning
> that this is a risk that the designers of PSK (external) intentionally
> ruled out. Going that path, it would be a step back in PSK security for
> TLS1.3.
>

Nikos,

Just as a clarification, I believe it's possible for a PSK-only
implementation to
avoid this attack by behaving uniformly for "unknown ID" and "invalid
binder"

-Ekr


> regards,
> Nikos
>
>