| Internet Engineering Task Force | S.B. Barbato |
| Internet-Draft | S.D. Dorigotti |
| Intended status: Informational | T.F. Fossati, Ed. |
| Expires: August 25, 2011 | KoanLogic |
| February 21, 2011 |
SCS: Secure Cookie Sessions for HTTP
draft-secure-cookie-session-protocol-00
This document provides an overview of SCS, a cryptographic protocol layered on top of the HTTP cookie facility, that allows an origin server to handle session state without storing it locally.
Its typical use cases include devices with little or no storage offering some functionality via an HTTP interface, and web applications with High Availability or load balancing requirements which may want to handle application state without the need to synchronize the pool through shared storage or peering.
Nevertheless, its security properties allow it to be used whenever privacy and integrity of cookies is a concern, at the cost of increased server CPU and bandwidth usage, and of some "credential-ownership" implications which will be thoroughly analysed.
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet- Drafts is at http://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on August 25, 2011.
Copyright (c) 2011 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
SCS is a cryptographic protocol layered on top of the HTTP cookie facility [I-D.ietf-httpstate-cookie], that allows an origin server to handle clients' session state without storing it locally.
An SCS enabled server delegates the application state storage to the client (e.g. a browser) - which basically acts as a remote storage device. A set of cryptographic transformations is used to ensure that information authenticity and confidentiality attributes of state data have the same characteristics as for typical "server-side" sessions.
Anyway, a peculiar difference between SCS and "server-side" cookie sessions arises when we carefully consider the roles of the playing entities. In the "server-side" model, the Server acts a triple role as the "generator", the "owner", and the "verifier" of cookie credentials. Instead, a server implementing SCS acts the "generator" and "verifier" roles only -- the "owner" being inapplicable as long as we have imposed the no-storage requirement.
In all respects, the Server grants the custody of the generated cookie to the Client, whose trust model needs to be taken into consideration when designing applications using SCS. The consequences of such discrepancy (e.g. deliberate deletion of a cookie, explicit privilege revocation, etc.) will be explored and analyzed in Section 7.2.
The no-storage requirement, which is the key design constraint of SCS, makes it an ideal candidate in the following settings:
An SCS server can be implemented within a web application by means of a user library that exposes the core SCS functionality and leaves explicit control over SCS cookies to the programmer, or transparently, by hiding the "diskless session" facility behind a generic session API abstraction.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in [RFC2119].
The SCS protocol defines:
Note that the PDU is transmitted to the client as an opaque data block, hence no interpretation nor validation is necessary.
The single requirement for client-side support of SCS is cookie activation on the browser. Only the server is involved in the PDU manipulation process.
In the following sections we assume S to be one or more interchangeable HTTP server entities (e.g. a server pool in a load-balanced or HA environment) and C to be the client with a cookie-enabled browser, or User Agent with equivalent capabilities.
S and C exchange the same PDU (Section 3.3), which consists of a set of interdependent cookies tied together by cryptographic transformations.
Confidentiality is limited to the application state information (i.e. SCS_DATA cookie), while integrity and authentication apply to the entire PDU.
The following subsections describe the syntax and semantics of of each SCS cookie.
Timestamp relating to the last read or write operation performed on session data, encoded a numeric string holding the number of seconds since UNIX epoch.
This value is updated with each client contact and is used to identify expired sessions.
If the received SCS_ATIME value is older than a predefined "session_max_age" (which is chosen by S as an application-level parameter), a session is considered to be no longer valid, and is therefore rejected.
Block of encrypted and optionally compressed data, containing session state. Note that no restriction is imposed on plain text structure: the protocol is completely agnostic as to state data layout.
If the total size of the SCS_DATA cookie, including name, value and attributes, exceeds 4096 bytes (see Section 6.1. of [I-D.ietf-httpstate-cookie]), it is sliced into n SCS_DATA{n} cookies, each 4KB in size, so that the concatenation of their values ordered by cookie name (1, 2, ..., N) yields the original SCS_DATA.
It is suggested [I-D.ietf-httpstate-cookie] that browsers accept at least 50 cookies per domain, which could lead to a theoretical limit of 184 KB as the maximum allowed state data block.
Anyway, in order to minimize both network bandwidth and client cookie store consumption, applications should try to upper bound state data to some sensible value. Also, an SCS implementation MAY decide to limit the accepted state data to any value greater than or equal to 4KB.
ASCII string that uniquely identifies the transform set (keys and algorithms) used to generate this PDU.
SCS assumes that a key-agreement/distribution mechanism exists for environments in which S consists of multiple servers (it may consist of a simple key-refresh in the case |S|=1), which provides a unique external identifier for each transform set defined and shared amongst pool members.
This identifier (equivalent to a SPI in a Data Security SA [RFC3740]) is represented in the cookie value.
Initialization Vector used for the encryption algorithm (Section 3.2).
In order to avoid providing correlation information to a possible attacker with access to a sample of SCS PDUs, the IV MUST be created randomly for each PDU.
Authentication tag based on the concatenation of SCS_ATIME, SCS_DATA, SCS_TID and SCS_IV.
The concatenation operation is done by packing the four strings containing the base64 encoded values in order (e.g.: "ZGF0YQ==YXRpbWU=dGlkaXY="), and supplying the resulting string to HMAC.
SCS could potentially use any combination of primitives capable of performing authenticated encryption. In practice an encrypt-then-mac approach [Kohno] with CBC-mode encryption and HMAC [RFC2104] authentication was chosen.
The two algorithms MUST be associated with two independent keys.
The possibility of using UMAC for authentication [RFC4418] has been taken into consideration, but priority was given to space over performance: the nonce field transfer would require an extra cookie, therefore reducing the space reserved to state information by another 4KB.
The following conventions will be used in the algorithm description (Section 3.2.4 and Section 3.2.5):
Implementors MUST support at least the following algorithms: [Bellare], are widely available, and can be implemented in a few kilobytes of memory, providing an extremely valuable feature in constrained devices.
which appear to be sufficiently secure in a wide range of use cases
One should consider using larger cryptographic key lengths (192 or 256 bit) according to the actual security and overall system performance requirements.
Compression, which may be useful or even necessary when handling large quantities of data, is not compulsory (in such case Comp/Uncomp are replaced by an identity matrix). If this function is enabled, DEFLATE [RFC1951] format MUST be supported.
Compression should not be enabled when handling relatively short and entropic state (e.g. pseudo random session identifiers). Instead, large and quite regular state blobs could get a significant boost when compressed.
Base-64 [RFC4648] is used for encoding/decoding of SCS cookie values. It is very wide-spread, and falls smoothly into the encoding rules defined in Section 4.1.1 of [I-D.ietf-httpstate-cookie].
The output data transformation as seen by the server (the only actor which manipulates PDUs) is illustrated by the pseudo-code in Figure 1.
1. iv = RAND()
2. atime = NOW
3. data = Enc(Comp(state))
4. tag = HMAC(e(data)||e(atime)||e(tid)||e(iv))
NOW is defined as the current timestamp of the server clock.
Since the only user of the atime field is the server, it is unnecessary for it to be synchronized with the client. However, if multiple servers are active in a load-balancing configuration, clocks SHOULD be synchronized to avoid errors in the calculation of session expiry.
If the length of (compressed) state is not a multiple of the block size, its value MUST be filled with padding bytes of equal value as the pad length -- equivalent to the scheme defined in Section 6.3 of [RFC5652].
Hence the SCS PDU fields are created as follows:
The inbound transformation is described in Figure 2.
1. If (tid is available)
2. data' = d($SCS_DATA)
atime' = d($SCS_ATIME)
tid' = d($SCS_TID)
iv' = d($SCS_IV)
tag' = d($SCS_AUTHTAG)
3. tag = HMAC(<data'>||<atime'>||<tid'>||<iv'>)
4. If (tag == tag' && NOW - atime' <= session_max_age)
5. state = Uncomp(Dec(data'))
6. Else discard PDU
7. Else discard PDU
If the cryptographic credentials (encryption and authentication algorithms and keys identified by SCS_TID) are unavailable (step 7.), the inbound PDU cannot be interpreted correctly.
This may happen for several reasons: e.g., if a device without storage has been reset and loses the credentials stored in RAM, if a server pool node desynchronizes, or in case of a key compromise that forces the invalidation of all available TIDs, etc.
Note that step 4. allows any altered packets or expired sessions to be discarded, hence avoiding unnecessary state decryption and decompression.
SCS can be modeled in the same manner as a typical store-and-forward protocol, in which the endpoints are S, consisting of one or more HTTP servers, and the client C, an intermediate node used to "temporarily" store the data to be successively forwarded to S.
In brief, S and C exchange an immutable cookie data block (Section 3.1): the state is stored on the client at the first hop and then restored on the server at the second, as in Figure 3.
1. dump-state:
Set-Cookie: SCS_DATA=...;
Expires=...; Path=...;
Domain=...;
S --> --> C
Set-Cookie: SCS_TID=...;
Expires=...; Path=...;
Domain=...;
...
2. restore-state:
Cookie: SCS_DATA=...;
C --> Cookie: SCS_TID=...; --> S
...
Note that although SCS cookies always have the same naming, there can be multiple active SCS sessions in use at a given user-agent as long as the tuple <SCS PDU, Domain, Path> is different.
SCS cookies MUST NOT be folded into a single HTTP header field, see Section 3 of [I-D.ietf-httpstate-cookie].
All SCS cookies belonging to the same PDU MUST carry the same attributes' set. This is not elegant nor bandwidth-friendly solution, but it is necessary in order to guarantee PDU coherence.
In the following sub paragraphs a series of recommendations is provided in order to maximize SCS PDU fitness in the generic cookie ecosystem.
SCS cookies MUST include an Expires attribute which shall be set to a value consistent with session_max_age.
For maximum compatibility with existing user agents the timestamp value MUST be encoded in rfc1123-date format which requires a 4-digit year.
Since not all UAs support this attribute, it MUST NOT be present in any SCS cookie.
SCS cookies MUST include a Domain attribute compatible with application usage.
A trailing '.' MUST NOT be present in order to minimize the possibility of a user-agent ignoring the attribute value.
This attribute MUST always be asserted when SCS sessions are carried over a TLS channel.
This specification provides some common recommendations and praxis relevant to cryptographic key management.
In the following, the term 'key' references both encryption and HMAC keys.
Furthermore, to preserve the validity of active HTTP sessions upon renewal of cryptographic credentials (whenever the value of SCS_TID changes), an SCS server MUST be capable of managing at least two transforms contemporarily: the currently instantiated one, and its predecessor.
Each transform set SHOULD be associated with an attribute pair: "refresh" and "expiry", which is used to identify the exposure limits (in terms of time or quantity of encrypted and/or authenticated bytes, etc) of related cryptographic material.
In particular, the "refresh" attribute specifies the time limit for substitution of transform set T with new material T'. From that moment onwards, and for an amount of time determined by "expiry", all new sessions will be created using T', while the active T-protected ones go through a translation phase in which:
T' {not valid yet} |---------------------|----------------
| translation stage |
T ----------------|---------------------| {no longer valid}
refresh refresh + expiry
As shown in Figure 4, the duration of the HTTP session MUST fit within the lifetime of a given transform set (i.e. from creation time until "refresh" + "expiry").
In practice, this should not be an obstacle because the longevity of the two entities (HTTP session and SCS transform set) should differ by one or two orders of magnitude.
An SCS server may take this into account by determining the duration of a session adaptively according to the expected deletion time of the active T, or by setting the "expiry" value to at least the maximum lifetime allowed by an HTTP session.
Since there is only one refresh attribute also in situations with more than one key (e.g. one for encryption and one for authentication) within the same T, the smallest value is chosen.
We would like to thank David Wagner and Lorenzo Cavallaro for their valuable feedback on this document.
This memo includes no request to IANA.
From a cryptographic architecture perspective, the described mechanism can be easily traced to an Encode-then-EtM scheme described in [Kohno].
Given a "provably-secure" encryption scheme and MAC (as for the algorithms mandated in Section 3.2.1), Kohno et al. [Kohno] demonstrate that their composition results in a secure authenticated encryption scheme.
The fact that the server does not own the cookie it produces, gives rise to a series of consequences that must be clearly understood when one envisages the use of SCS as a cookie provider and validator for his/her application.
In the following paragraphs, a set of different attack scenarios (together with corresponding countermeasures where applicable) are identified and analyzed.
SCS doesn't address replay of old cookie values.
In fact, there is nothing that guarantees an SCS application about the client having returned the most recent version of the cookie.
As with "server-side" sessions, if an attacker gains possession of a given user's cookies - via simple passive interception or another technique - he/she will always be able to restore the state of an intercepted session by representing the captured data to the server.
The SCS_ATIME value along with the session_max_age configuration parameter allow SCS to mitigate the chances of an attack (by forcing a time window outside of which a given cookie is no longer valid), but cannot exclude it completely.
A countermeasure against the "passive interception and replay" scenario can be applied at transport/network level using the anti-replay services provided by e.g., SSL/TLS [RFC5246] or IPsec [RFC4301].
Anyway, a generic solution is still out of scope: an SCS application wishing to be replay-resistant must put in place some ad hoc mechanism to prevent clients (both rogue and legitimate) from (a) being able to replay old cookies as valid credentials and/or (b) getting any advantage by replaying them.
In the following, some typical use cases are illustrated:
It may be noteworthy that despite the chances of preventing replay in some well established circumstances by using aforementioned mechanisms, if the attacker is able to use the cookie before the legitimate client gets a chance to, then the impersonation attack will always succeed.
A direct, and important, consequence of the missing owner role in SCS is that a client could intentionally delete its cookie and return nothing.
The application protocol has to be designed so there is no incentive to do so, for instance:
Note that this behavior is not equivalent to cookie removal in the "server-side" cookie model, because in case of missing cookie backup by other parties (e.g. the application using SCS), the Client could simply make it disappear once and for all.
SCS doesn't prevent sharing (both voluntary and illegitimate) of cookies between multiple clients.
In the context of voluntary cookie sharing, using HTTPS is useless: Client certificates are just as shareable as cookies, hence equivalently to the "server-side" cookie model, there seems to be no way to prevent this threat.
The theft could be mitigated by securing the wire (e.g. via HTTPS, IPsec, VPN, ...), thus reducing the opportunity of cookie stealing to a successful attack on the protocol endpoints.
Note that all the abovementioned vulnerabilities also apply to typical server-side sessions, making SCS at least as secure (based on the current analysis), but there are a few good reasons to consider its security level enhanced.
First of all, the confidentiality feature provided by SCS protects cookie state information which is normally plain-text.
Furthermore, none of the common vulnerabilities of server-side sessions (SID prediction, SID brute forcing, session fixation [Kolsec]) can be exploited when using SCS, unless the attacker possesses encryption and HMAC keys (both current ones and those relating to the previous set of credentials).
More generally no slicing nor altering operations can be done over an SCS PDU without controlling the cryptographic keyset and cipherset.
| [RFC2119] | Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, March 1997. |
| [RFC4648] | Josefsson, S., "The Base16, Base32, and Base64 Data Encodings", RFC 4648, October 2006. |
| [RFC1951] | Deutsch, P., "DEFLATE Compressed Data Format Specification version 1.3", RFC 1951, May 1996. |
| [RFC4086] | Eastlake, D., Schiller, J. and S. Crocker, "Randomness Requirements for Security", BCP 106, RFC 4086, June 2005. |
| [RFC2104] | Krawczyk, H., Bellare, M. and R. Canetti, "HMAC: Keyed-Hashing for Message Authentication", RFC 2104, February 1997. |
| [RFC5652] | Housley, R., "Cryptographic Message Syntax (CMS)", STD 70, RFC 5652, September 2009. |
| [I-D.ietf-httpstate-cookie] | Barth, A, "HTTP State Management Mechanism", Internet-Draft draft-ietf-httpstate-cookie-22, February 2011. |
| [RFC5246] | Dierks, T. and E. Rescorla, "The Transport Layer Security (TLS) Protocol Version 1.2", RFC 5246, August 2008. |
| [RFC3740] | Hardjono, T. and B. Weis, "The Multicast Group Security Architecture", RFC 3740, March 2004. |
| [RFC4418] | Krovetz, T., "UMAC: Message Authentication Code using Universal Hashing", RFC 4418, March 2006. |
| [RFC4301] | Kent, S. and K. Seo, "Security Architecture for the Internet Protocol", RFC 4301, December 2005. |
| [Bellare] | Bellare, M.B., "New Proofs for NMAC and HMAC: Security Without Collision-Resistance", 2006. |
| [Kohno] | Kohno, T.K., Palacio, A.P. and J.B. Black, "Building Secure Cryptographic Transforms, or How to Encrypt and MAC", 2003. |
| [Kolsec] | Kolsec, M.K., "Session Fixation Vulnerability in Web-based Applications", 2002. |
A reference implementation (at present in very early stage) of the SCS protocol can be found at http://github.com/koanlogic/libscs.