Eric Rescorla
2018-03-08 02:58:42 UTC
Eric Rescorla has entered the following ballot position for
draft-ietf-core-object-security-09: 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 IESG DISCUSS and COMMENT positions.
The document, along with other ballot positions, can be found here:
https://datatracker.ietf.org/doc/draft-ietf-core-object-security/
----------------------------------------------------------------------
DISCUSS:
----------------------------------------------------------------------
https://mozphab-ietf.devsvcdev.mozaws.net/D3075
DISCUSS
My overall concern with this document is that I am unable to evaluate
the security properties of the system. I have described a number of
issues below, but the basic problem is that this sort of partial
protection is extremely hard to reason about and the security
considerations do not do an adequate job of evaluating the impact of
proxies modifying these values. I am similarly concerned about the
HTTP mapping and link section which seems extremely sketchy and has
essentially no security analysis, and yet potentially have a lot
of landmines.
At minimum, this document needs to walk through the implications
of modifications by the proxy to every unprotected field in
the pure CoAP context as well as the HTTP context (if you want
to retain that binding).
are given in Appendix A. OSCORE does not depend on underlying
layers, and can be used anywhere where CoAP or HTTP can be used,
including non-IP transports (e.g., [I-D.bormann-6lo-coap-802-15-ie]).
IMPORTANT: This document claims to be applicable to protocols other
than COAP, in particular HTTP. Has this been reviewed by the HTTP
working group? Martin Thomson's review suggests that this is out of
step with HTTP practice.
IDs MUST be long uniformly random distributed byte strings such that
the probability of collisions is negligible.
IMPORTANT: I don't understand how this paragraph and the previous
paragraph interact. You say that the maximum length is 7 octets in the
previous paragraph, which I don't think qualifies as "long".
| 1 | If-Match | x | |
| 3 | Uri-Host | | x |
| 4 | ETag | x | |
IMPORTANT: Why do you think that it's not important to have integrity
protection for Uri-Host and Uri-port?
Outer option message fields (Class U or I) are used to support proxy
operations.
IMPORTANT: This seems problematic because the proxy cannot verify class I
fields.
layer only, and not the Messaging Layer (Section 2 of [RFC7252]), so
fields such as Type and Message ID are not protected with OSCORE.
IMPORTANT: This seems extremely hard to reason about. What are the
implications of the proxy being able to change these?
o request_piv: contains the value of the 'Partial IV' in the COSE
object of the request (see Section 5).
IMPORTANT: I think what I am getting here is that the request_piv is
used to verify that the request and response match. However, I do not
see this explicitly stated anywhere, and it's not clear to me how the
client is supposed to recover the request_piv and the text is pretty
unclear here? Is the external_aad carried somewhere in the message? Am
I supposed to reconstruct it from the message id?
For responses, the message binding guarantees that a response is not
older than its request. For responses without Observe, this gives
IMPORTANT: I am not sure that this is true. What happens of the
counterparty lies? What is your threat model?
An extension of OSCORE may also be used to protect group
communication for CoAP [I-D.tiloca-core-multicast-oscoap]. The use
of OSCORE does not affect the URI scheme and OSCORE can therefore be
used with any URI scheme defined for CoAP or HTTP. The application
decides the conditions for which OSCORE is required.
This is pretty surprising to just drop in here. Multicast has totally different
security properties from non-multicast.
----------------------------------------------------------------------
COMMENT:
----------------------------------------------------------------------
but is also able to eavesdrop on, or manipulate any part of the
message payload and metadata, in transit between the endpoints. The
proxy can also inject, delete, or reorder packets since they are no
Nit: you want
"eavesdrop on, or manipulate any part of, the message payload and metadata in
transit"
I.e., move the second comma
the endpoints, and those are therefore processed as defined in
[RFC7252]. Additionally, since the message formats for CoAP over
unreliable transport [RFC7252] and for CoAP over reliable transport
Nit: "OSCORE protects neither .... nor...."
Salt. Length is determined by the AEAD Algorithm. Its value is
immutable,
different operations. One mechanism enabling this is specified in
[I-D.ietf-core-echo-request-tag].
Is this a security condition?
of [RFC7252], where the delta is the difference to the previously
included class I option.
Is the delta here the previously included Class I option or the previously
included instance of the same option, as it appears to say in S 3.1.
compressed COSE object. The values n = 6 and n = 7 are
reserved.
How can Partial IV not be present? it's the sequence number. Is the answer that
it is the 0 value?
response. The server therefore needs to store the kid and Partial IV
of the request until all responses have been sent.
It was my understanding that the kid was needed to look up the key. Why are kid
substitution attacks an issue?
The maximum Sender Sequence Number is algorithm dependent (see
Section 11), and no greater than 2^40 - 1. If the Sender Sequence
Number exceeds the maximum, the endpoint MUST NOT process any more
If you take my suggestion about removing senderID from the nonce you will be
able to relax this.
After boot, an endpoint MAY reject to use existing security contexts
from before it booted and MAY establish a new security context with
Nit: this is ungrammatical
included in the message. If the AEAD nonce from the request was
used, the Partial IV MUST NOT be included in the message.
IMPORTANT: You are now violating the invariant of using the same nonce twice.
That's fine in this case, because you have per-sender keys but it demonstrates
that it is unnecessary to encode the sender_id in the nonce field.
Security level here means that an attacker can recover one of the m
keys with complexity 2^(k + n) / m. Protection against such attacks
can be provided by increasing the size of the keys or the entropy of
This paragraph is extremely hard to follow but I am not persuaded that it is
correct. Do you have a citation for the claim that you can add the key entropy
and the nonce entropy like this.
style of padding means that all values need to be padded. Similar
arguments apply to other message fields such as resource names.
The PKCS#7 padding scheme at minimum has potential timing channels
The server verifies that the Partial IV has not been received before.
The client verifies that the response is bound to the request.
How does the client verify this
Partial IV (in network byte order) with zeroes to exactly nonce
length - 6 bytes,
IMPORTANT: I don't understand the reason for this construction. You
already require that the key be derived via HKDF from the |master key|
and |sender ID| therefore, it is not necessarily to separately encode
the sender ID in the nonce. This would ordinarily not be a large
issue, but as it requires very extreme restrictions on the sender ID
(and essentially precludes random sender IDs) I believe it is worth
considering changing, but it's ultimately a WG decision, hence not
in my discuss.
draft-ietf-core-object-security-09: 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 IESG DISCUSS and COMMENT positions.
The document, along with other ballot positions, can be found here:
https://datatracker.ietf.org/doc/draft-ietf-core-object-security/
----------------------------------------------------------------------
DISCUSS:
----------------------------------------------------------------------
https://mozphab-ietf.devsvcdev.mozaws.net/D3075
DISCUSS
My overall concern with this document is that I am unable to evaluate
the security properties of the system. I have described a number of
issues below, but the basic problem is that this sort of partial
protection is extremely hard to reason about and the security
considerations do not do an adequate job of evaluating the impact of
proxies modifying these values. I am similarly concerned about the
HTTP mapping and link section which seems extremely sketchy and has
essentially no security analysis, and yet potentially have a lot
of landmines.
At minimum, this document needs to walk through the implications
of modifications by the proxy to every unprotected field in
the pure CoAP context as well as the HTTP context (if you want
to retain that binding).
are given in Appendix A. OSCORE does not depend on underlying
layers, and can be used anywhere where CoAP or HTTP can be used,
including non-IP transports (e.g., [I-D.bormann-6lo-coap-802-15-ie]).
IMPORTANT: This document claims to be applicable to protocols other
than COAP, in particular HTTP. Has this been reviewed by the HTTP
working group? Martin Thomson's review suggests that this is out of
step with HTTP practice.
IDs MUST be long uniformly random distributed byte strings such that
the probability of collisions is negligible.
IMPORTANT: I don't understand how this paragraph and the previous
paragraph interact. You say that the maximum length is 7 octets in the
previous paragraph, which I don't think qualifies as "long".
| 1 | If-Match | x | |
| 3 | Uri-Host | | x |
| 4 | ETag | x | |
IMPORTANT: Why do you think that it's not important to have integrity
protection for Uri-Host and Uri-port?
Outer option message fields (Class U or I) are used to support proxy
operations.
IMPORTANT: This seems problematic because the proxy cannot verify class I
fields.
layer only, and not the Messaging Layer (Section 2 of [RFC7252]), so
fields such as Type and Message ID are not protected with OSCORE.
IMPORTANT: This seems extremely hard to reason about. What are the
implications of the proxy being able to change these?
o request_piv: contains the value of the 'Partial IV' in the COSE
object of the request (see Section 5).
IMPORTANT: I think what I am getting here is that the request_piv is
used to verify that the request and response match. However, I do not
see this explicitly stated anywhere, and it's not clear to me how the
client is supposed to recover the request_piv and the text is pretty
unclear here? Is the external_aad carried somewhere in the message? Am
I supposed to reconstruct it from the message id?
For responses, the message binding guarantees that a response is not
older than its request. For responses without Observe, this gives
IMPORTANT: I am not sure that this is true. What happens of the
counterparty lies? What is your threat model?
An extension of OSCORE may also be used to protect group
communication for CoAP [I-D.tiloca-core-multicast-oscoap]. The use
of OSCORE does not affect the URI scheme and OSCORE can therefore be
used with any URI scheme defined for CoAP or HTTP. The application
decides the conditions for which OSCORE is required.
This is pretty surprising to just drop in here. Multicast has totally different
security properties from non-multicast.
----------------------------------------------------------------------
COMMENT:
----------------------------------------------------------------------
but is also able to eavesdrop on, or manipulate any part of the
message payload and metadata, in transit between the endpoints. The
proxy can also inject, delete, or reorder packets since they are no
Nit: you want
"eavesdrop on, or manipulate any part of, the message payload and metadata in
transit"
I.e., move the second comma
the endpoints, and those are therefore processed as defined in
[RFC7252]. Additionally, since the message formats for CoAP over
unreliable transport [RFC7252] and for CoAP over reliable transport
Nit: "OSCORE protects neither .... nor...."
Salt. Length is determined by the AEAD Algorithm. Its value is
immutable once the security context is established.
Nit: you might just say above or below this list that all the values areimmutable,
different operations. One mechanism enabling this is specified in
[I-D.ietf-core-echo-request-tag].
Is this a security condition?
of [RFC7252], where the delta is the difference to the previously
included class I option.
Is the delta here the previously included Class I option or the previously
included instance of the same option, as it appears to say in S 3.1.
compressed COSE object. The values n = 6 and n = 7 are
reserved.
How can Partial IV not be present? it's the sequence number. Is the answer that
it is the 0 value?
response. The server therefore needs to store the kid and Partial IV
of the request until all responses have been sent.
It was my understanding that the kid was needed to look up the key. Why are kid
substitution attacks an issue?
The maximum Sender Sequence Number is algorithm dependent (see
Section 11), and no greater than 2^40 - 1. If the Sender Sequence
Number exceeds the maximum, the endpoint MUST NOT process any more
If you take my suggestion about removing senderID from the nonce you will be
able to relax this.
After boot, an endpoint MAY reject to use existing security contexts
from before it booted and MAY establish a new security context with
Nit: this is ungrammatical
included in the message. If the AEAD nonce from the request was
used, the Partial IV MUST NOT be included in the message.
IMPORTANT: You are now violating the invariant of using the same nonce twice.
That's fine in this case, because you have per-sender keys but it demonstrates
that it is unnecessary to encode the sender_id in the nonce field.
Security level here means that an attacker can recover one of the m
keys with complexity 2^(k + n) / m. Protection against such attacks
can be provided by increasing the size of the keys or the entropy of
This paragraph is extremely hard to follow but I am not persuaded that it is
correct. Do you have a citation for the claim that you can add the key entropy
and the nonce entropy like this.
style of padding means that all values need to be padded. Similar
arguments apply to other message fields such as resource names.
The PKCS#7 padding scheme at minimum has potential timing channels
The server verifies that the Partial IV has not been received before.
The client verifies that the response is bound to the request.
How does the client verify this
Partial IV (in network byte order) with zeroes to exactly nonce
length - 6 bytes,
IMPORTANT: I don't understand the reason for this construction. You
already require that the key be derived via HKDF from the |master key|
and |sender ID| therefore, it is not necessarily to separately encode
the sender ID in the nonce. This would ordinarily not be a large
issue, but as it requires very extreme restrictions on the sender ID
(and essentially precludes random sender IDs) I believe it is worth
considering changing, but it's ultimately a WG decision, hence not
in my discuss.