As has been discussed on the mailing list and in several recent WG
meetings, even with TLS (which I'm proposing we drop), we do not
currently have effective replay protection in the RPKI up-down
protocol (expired draft-ietf-sidr-rescerts-provisioning).  The
solutions for this are relatively straightforward, we just have to
update the spec to include them.

The first level of replay protection would be just to check the
timestamp in the CMS message, and enforce a rule that any new message
from a particular peer (protocol, not routing sense of "peer") must
have a newer timestamp than any we've seen until now.   The CMS
profile we're using already includes timestamps, so we get this for
free, and we don't need synchronized clocks, we just need to know that
the timestamp as viewed by the sending party has increased.  This, by
itself, should be enough to detect replay attempts.

The second level of protection would be if we also need to detect
messages being dropped out of the protocol exchange (ie, this one is
about selective message blocking, not replay per se).  A timestamp
won't suffice for this, but a counter will.  So if we're concerned
about this we should add a serial number to the XML header.  Geoff
proposed something like this back in 2007, although his proposed
mechanism was a bit more complicated than I think we'd really need for
this threat.  All we'd really need would be an XML attribute
containing a simple nonnegative integer serial number: the client
would increment this to the next value when sending a new message, the
server would just echo back the client's value.

I'm reasonably certain that we need the first rule (timestamp check),
as there is at least one known attack (replay of an old issuance
request) without it.  I'm not sure we need the serial number, but
Steve Bellovin recommended that we at least think about it.

For both of these, there is the little issue of what an implementation
is meant to do if the rule is violated.  Log something, almost
certainly, but should it reject the message?  Probably, for timestamp
violation, similarly for serial number repeating or going backwards.
Not quite so obvious for a gap in the serial number sequence, seems
like rejecting the message in this case would invite a cheap denial of
service attack, so my inclination would be to raise an alert but allow
the message in this case.

At minimum, I propose that we add the timestamp check rule.  This
requires no change to the message format.
_______________________________________________
sidr mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/sidr

Reply via email to