On Tue, Mar 10, 2026 at 11:52:29PM +0000, John Mattsson wrote:
> Ilari Liusvaara wrote:
> >E.g., a new mechanism that can change client identity would be insecure
> >with HTTP/2 and HTTP/3.
> 
> I agree that none of the mechanisms should allow changes to the client
> or server identity.
> 
> Ilari Liusvaara wrote:
> >What I meant is that there does not seem to be much simpler way to do
> >what EKU does.
> 
> A major problem with the current EKU draft is that mutual
> reauthentication requires changes to the application layer. And if
> application-layer modifications are possible, it is simpler to
> establish subsequent overlapping connections, similar to
> reauthentication in IPsec.
> https://docs.strongswan.org/docs/latest/config/rekeying.html

It could be possible to add signatures from both peers using
previous long-term key (if any).

The only situation I can come up with where that would help is attacker
bleeding the session keys, and then becoming a MITM, rewriting any
rekeys.

However, if client does not have long-term key, that does not prevent
attacker from using keys bled from either endpoint for hijacking the
connection.

I suppose it would be technically possible for client to send a
temporary key for authenticating key updates if it does not have a
long-term key.


> >As to why I think the threat model is dubious: It seems to be aimed at
> >risk of old secrets being exposed, which I think is very rare. Re-
> >authentication would at least help against bleeding-type attacks, but
> >AFAIK those are rare compared to RCE-type attacks (or compromising the
> >whole TCB).
> >
> >As to keeping long-term key in TEE, but application traffic secrets
> >outside it... Keeping application traffic secrets in TEE is not
> >feasible, due to TEE limitations and latency.
> >
> >It is also not feasible to use hardware memory protection, because the
> >memory overheads are prohibitive (even without the current RAM
> >shortage).
> 
> My view on the threat model:
> 
> It is common to protect the signature key more strongly than the
> session keys. This can be achieved in several ways, such as using
> an HSM, TPM, TEE, or privilege separation. Therefore, it is reasonable
> to assume that session keys might be compromised while the signature
> key remains secure.

Yes, doing that is not hard.


> I assume that an MLS implementation could, and maybe should,
> protect the private KEM keys more strongly than the session
> keys. I think this requires security considerations in the
> drafts.

The private KEM keys live for very short period anyway.

And details on how rekeying works are actually relevant here.
Protecting the KEM keys requires four-message rekey, it does not work
with three-message rekey.

But protecting session keys is not an option due to the existing
rekey mechanism in RFC8446.


> draft-tian-quic-quicmls-00 states:
> >MLS provides authenticated key agreement
> 
> My understanding is that for the initial shared secret established
> via KeyPackage and Welcome, MLS provides key transport rather than
> key agreement. MLS in TLS/QUIC provides key agreement for the derived
> secrets. I think the drafts should clarify this distinction and
> explain that ClientHello.random remains important for ensuring the
> intended security properties.

Briefly looking at draft-tian-quic-quicmls-00, it looks to work in
somewhat different way from draft-housley-tls-using-mls-handshake.
I still do not get how draft-housley-tls-using-mls-handshake is
supposed to work (even ignoring any negotiation issues).

And yes, MLS provides key transport, not key agreement. 




-Ilari

_______________________________________________
TLS mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to