SMTorP
SMTorP is the Simple Mail Tor Protocol (a.k.a. SMTP over Tor).
The idea can be summarized as follows:
- SMTorP is e-mail, but delivered over a secure channel which protects the metadata as well as message contents.
- SMTorP defines an e-mail address format: addresses ending in .onion
- SMTorP servers are enhanced SMTP servers which run as Tor hidden services
- Mailpile as an MUA should ship with both an SMTorP client and server built-in, and be bundled with Tor (Win, Mac) or depend on Tor on Linux.
Note: This document is a rough draft of notes and ideas, once this has been polished and proof of concept code written, we should draft an RFC and get the community to weigh in.
A very rough 1st proof of concept was added to Mailpile at the Tor developer summit hackdays (Reykjavík, Feb 20th & 21st, 2014) validating that the idea is at least quite simple to implement. To test it yourself:
-
Install Tor and the Python socksipy or socksipychain module
-
Configure Tor for a hidden service mapping
something.onion:25tolocalhost:33412 -
Enable the smtp_server plugin:
mailpile> append sys.plugins smtp_server
-
Restart Mailpile.
-
Set the SMTP server port to the port used above:
mailpile> set sys.smtpd.port = 33412
-
Restart Mailpile again.
You should now be able to send mail to foo@bar.onion, and receive
incoming on the Tor hidden service domain configured in step 2.
WARNING: This exposes the built-in SMTP server to the Tor network, anyone who knows (or guesses?) your .onion address may connect to the SMTP server and wreak havoc. The SMTP server is currently known to be a security hole, see https://github.com/pagekite/Mailpile/issues/532 for details and progress on this issue.
The primary goal of SMTorP is to provide a fully decentralized and largely backwards-compatible upgrade path for e-mail, which aims to make it very difficult for an adversary to know that two users have communicated. The protocol should protect both message content and the metadata of e-mail from spying.
- SMTorP does not aim provide stronger security guarantees than the Tor network itself, which means that at the very least a sufficiently powerful adversary may be able to infer that two users are probably communicating, based on statistical analysis of network traffic at both endpoints (timing attacks). This is a problem inherited from Tor itself and is considered out of scope for SMTorP.
- Anonymous communication (allowing 2 parties to communicate securely without knowing each others identity) is not a primary goal for SMTorP, although users who take special care not to reveal personally identifying information may well be able to use it in this manner.
We are interested in implementing a mandatory hashcash-like proof of work system, to make bulk unsolicited e-mail over SMTorP economically infeasable.
Since hashcash was invented, hardware advances (primarily widespread GPU availability) have rendered the original proposals obsolete. However, research is being done into developing hashing algorithms which are "SIMD-hard", "Memory-hard" or difficult in other dimensions which may make them potential candidates. We need to look at the relevant literature (Wikipedia probably is probably a good place to start, if not Ella tells us to talk to Matt Green).
We also probably need to specify an upgrade path, so broken algorithms can be phased out and replaced with harder problems to solve. One thing to do here, is offer multiple algorithms, but scale the difficulty on deprecated algorithms faster than on the algos we actually like.
The SMTorP message format is the same as SMTP e-mail (RFC822 and successors). PGP/MIME is recommended.
The SMTorP server is a standard ESMTP server, exposed to the Internet using TLS on a Tor hidden service. TLS is mandatory (probably a minimum of version 1.2).
Open questions:
- Does SMTorP need additional protocol verbs for implementing hashcash or checking the state of relayed mail (when using an outgoing relay)?
- Would ever want to use STARTTLS? Why?
- Should we provide a stand-alone server implemetation, or just Mailpile native?
- How to manage keys and prevent MITM? TOFU? Fingerprints in addresses?
An SMTorP address has the same format as a normal SMTP e-mail address, with additional constraints on the domain part:
<user>@[<tls-cert-fingerprint>.]<tor-hidden-service-hash>.onion
This syntax has the benefit of being largely backwards compatible with existing MUAs and in line with user expectations, allowing a staged rollout of SMTorP.
The optional(?) tls-cert-fingerprint part of the address can be used to signal to the sending MUA what TLS credentials to expect the server to present upon connection. This improves the security of self-signed certificates and makes active man-in-the-middle attacks much more difficult. The difficulty level (and security) can be tuned by adjusting how many bits of the fingerprint are included in the e-mail address. MUAs could be encouraged to "upgrade" the addresses stored in a user's address book to the full length automatically on first use, the abbreviated form would primarily be manual scenarios.
The main drawback of this address syntax is it is not very human readable, hashes are notoriously long and hard to write. Some potential mitigation strategies are discussed below.
Users with SMTorP addresses may choose to advertise them in their e-mail signatures or in a custom header, allowing compatible MUAs to recognize that a more secure path is available and opportunistically upgrading to the more secure protocol.
An SMTorP address fits in a reasonably sized QR-code, for printing on business cards.
The format of SMTorP addresses is a subset of normal SMTP e-mail addresses and the standard mailto: URL spec can be used, although there is risk of poor user experience when legacy e-mail clients attempt to send mail to an SMTorP address (see deployment strategy below).
DNS TXT records could be used to map individual addresses, or entire domains to SMTorP addresses. The pros and cons of this approach should be explored further.
Ideally, SMTorP will be natively supported by MUAs, and Mailpile will provide the initial proof-of-concept.
However, since SMTorP inherits most of its characteristics from legacy e-mail it should be easy to implement plugins or proxies which add SMTorP support to legacy infrastructure.
Any traditional e-mail provider or in-house mail server should be able to upgrade their SMTP submission servers to support SMTorP addresses in addition to legacy e-mail, if they so desire, merely by installing Tor and configuring the SMTP server to use a different policy for delivery of .onion addresses. One partial implementation for exim was discussed by Johannes Berg.
Whether SMTorP should always be purely peer-to-peer or whether it should allow for intermediate relays (analogous to fallback MX servers in the SMTP world) is an open question. There are two primary relay strategies available:
- Sending relays (an SMTP submission server)
- Receiving relays (an SMTP MX server)
SMTorP could support either or both, but as the system is expected to operate independently of DNS, the question remains how to encode a fallback strategy into the addresses themselves without everything becoming illegible (always delivering through a relay is just asking for MITM attacks, we prefer direct delivery whenever possible).
[TODO: Write more. Bjarni doesn't like relays, Ella contends they are necessary for users with intermittent networking. We met in the middle agreeing that sending relays may be a reasonable compromise.]
Running a hidden service of this nature makes it easier to track whether a particular user is online, as long as you know their SMTorP address.
Although this is an active attack, it may be possible to check that the hidden service is online, without the hidden service ever seeing traffic, which makes this undetectable by the user.
If this is considered a serious problem, then it argues in favour of running SMTorP relays.
When deployed as a peer-to-peer solution, SMTorP alone provides strong protection against network-based monitoring.
However, the security of Tor hidden services is not generally considered to be entirely future-proof and SMTorP cannot protect the integrity of data at rest, once the mail has been delivered. So a "belt and suspenders" approach of using PGP/MIME for messages delivered over SMTorP is still recommended.
The only hypothetical reason to not use PGP/MIME, would be to allow for plausible deniability. It is our opinion that this currently has no real world benefit.
-
When a message is being sent over SMTorP, the user should be made aware, e.g. by putting a lock or SMTorP logo on the send button.
-
Mixing SMTorP and legacy e-mail addresses in the same e-mail should be prevented by the MUA, both for security reasons and to prevent the usability nightmare of the legacy recipients being unable to reply-all to everyone.