TLS 1.2 Session Tickets | Hacker News Reader

TLS 1.2 Session Tickets

blog.filippo.io

183 pointsgtank48 comments

The article and the comments here lack a bit of context. At the time where tickets were starting to be deployed (around 2011), ECDH ciphers were not in use and PFS was therefore quite expensive (and therefore, almost never used). Around the same time, the move from 1024-bit certificates to 2048-bit certificates were making TLS difficult to deploy at scale for everyone except Google.

Session tickets were primarily a mechanism to help to reduce to load of TLS connections. The reduced latency is a happy side-effect. At the time, it was already well-known that this would negatively impact PFS (something I already mention in my own blog article about session tickets in 2011).

Moreover, many seem to build a scenario where session tickets are recovered by an attacker but not the private keys. Usually, session tickets are in memory only and private keys are persisted. This makes those tickets more unlikely to be recovered by anyone unless you get access to a running web server. In this case, you have far more serious trouble than those tickets (even if the attacker still cannot decode past recorded data).

Even today, TLS handshakes are quite expensive, both for the server and the clients. Using session tickets (with rotated keys) still outweigh the downsides of not using them. That's why every major site are using them.

So if like me you want to know how to disable session tickets in Apache httpd, you need to be running at least httpd 2.4.8 with OpenSSL 1.0.2. Then you can set:

  SSLOpenSSLConfCmd Options -SessionTicket

Alternatively, if you're running at least httpd 2.4.11 with OpenSSL 0.9.8f, you can set this instead:

  SSLSessionTickets off

https://httpd.apache.org/docs/2.4/mod/mod_ssl.html#sslsessio...

Also, if you're running RHEL/CentOS 7, you'll want to switch to the SCL version of Apache httpd (2.4.25) as the version in the base repositories (2.4.6) is too old to support either setting:

https://www.softwarecollections.org/en/scls/rhscl/httpd24/

Well, technically RHEL7 runs httpd-2.4.6-67, which is RedHat's 67th patched version of 2.4.6.

They backport a lot of stuff from later versions, don't see why they could not have done this (they haven't, I checked).

> An attacker with the STEK doesn't need to wait until session resumption is attempted. Session Tickets containing the current session keys are sent at the beginning of every connection that merely supports Session Tickets. In plaintext on the wire, ready to be decrypted with the STEK, fully bypassing Diffie-Hellman.

I found this flaw to be the scariest, by far. It means that a connection with "forward secrecy" gives up this property as soon as it begins.

Gives up the property before it begins even. But I completely agree with you. I would go as far to say a connection does not have PFS if you are sending the DH session key encrypted only by a [geographically replicated] static key before the session starts. You are literally destroying the entire purpose of the DH handshake in doing so.

I always assumed the STEK-encrypted session key was sent inside the established TLS PFS stream. To send it outside the stream is mind-numbingly insane!

Browsers should not report connections as having PFS if tickets are enabled in TLS 1.2. This is like NSA slide "PFS added and removed here :-)"

You're clearly not destroying the entire purpose of the DH handshake, because the STEK is rotated and so the time window of compromise is limited to by the rotation interval. Even with poorly rotated ticket keys, you're still better off with a PFS suite than without it --- without the DH, every client has an unlimited window of exposure if the server's private key is compromised.

True, it does not destroy the entire purpose if and when the STEK is securely and properly rotated. But to call this PFS stretches the truth in ways that boring crypto never does.

The number of real-world cases where PFS is reduced exactly to the non-PFS state by the existence of session tickets with fixed STEKs is not only non-zero, but rather I would venture a guess it is staggeringly large.

E.g. Configuring Session Tickets [1]

While Apache offers the SSLSessionTicketKeyFile directive to specify a key file that should contain 48 random bytes, it is recommended to not specify one at all. Apache will simply generate a random key on startup and use that to encrypt session tickets for as long as it is running.

The good thing about this is that the session ticket key will not touch persistent storage, the bad thing is that it will never be rotated. Generated once on startup it is only discarded when Apache restarts. For most of the servers out there that means they use the same key for months, if not years.

To provide forward secrecy we need to rotate the session ticket key about daily and current Apache versions provide no way of doing that. The only way to achieve that might be use a cron job to gracefully restart Apache daily to ensure a new key is generated. That does not sound like a real solution though and nothing ensures the old key is properly overridden.

...

Nginx, too, provides no way to automatically rotate keys. Reloading its configuration daily using a cron job might work but does not come close to a real solution either.

...

HAproxy does not allow configuring session ticket parameters. It implicitly supports this feature because OpenSSL enables it by default. HAproxy will thus always generate a session ticket key on startup and use it to encrypt tickets for the whole lifetime of the process.

[1] - https://timtaubert.de/blog/2014/11/the-sad-state-of-server-s...

Note that's from 2014 -- has it changed since then?

The 'P' in 'PFS' has always stretched the truth, which is why a lot of practitioners elide it.

Even if you never rotate STEKs and never reset systems to get a new STEK accidentally, the DH handshake is still providing forward secrecy value for clients who don't do session tickets.

Do such clients actually exist in mainstream usage? The one crypto bug which you are actually protected from by using IE on Win7!

I don't know, but the cost that session tickets are saving are almost entirely an externality to clients, and the privacy benefit of not honoring tickets isn't, so there would seem to be some value to a Chrome extension that busted that cache.

Yeah but it's not the client sending the resumption ticket but sending the capability in the ClientHello which exposes the session key under a possibly fixed key. So you can't fix this by busting the client cache I think. You need the capability flag off.

Even so could an active adversary turn on the capability in the ClientHello and still get a chance to see the STEK encrypted session key even against legacy (or intentionally ticket-disabled) clients? This I don't know - would a client just drop the ticket message and proceed or would it abort? But at least it would have to be an active attacker.

Yep, this got pointed out to me several times on Slack. Session tickets suck.

This is by design. The primary goal of session tickets is to resume a session without doing excessive crypto (no asymmetric cryptography).

If you're using OpenSSL, you can disable RFC 5077 Session Tickets via SSL_OP_NO_TICKET:

https://wiki.openssl.org/index.php/SSL/TLS_Client#Session_Ti...

Figuring out how to access that option from your library/language is an exercise for the reader.

Looking into this for Python. Python 3 supports sessions and the tickets may be disabled:

https://docs.python.org/3/library/ssl.html#ssl.SSLSession

https://docs.python.org/3/library/ssl.html#ssl.OP_NO_TICKET

For Python 2, I see no way to get at the session object (at least on the client side), much less disable session tickets:

https://docs.python.org/2/library/ssl.html

Under Python 2, pyOpenSSL has better support. You can make use of it on Python2 (and 3) via requests by installing `requests[security]` instead of just `requests`. Using `[security]` causes requests to pull in pyOpenSSL, cryptography and idna packages.

Under the covers, requests is using urllib3 and it ends up makes this call if pyOpenSSL is installed:

    urllib3.contrib.pyopenssl.inject_into_urllib3()

http://urllib3.readthedocs.io/en/latest/reference/urllib3.co...

That's as far as I've gotten. There's no documentation so it's going to require reading the urllib3 source to figure out what's going on under the hood.

Edit: nope, urllib3 doesn't support SSL session re-use. There's an open PR:

https://github.com/shazow/urllib3/issues/590

What is Cloud Flare's policy for managing STEKs? Are they distinct per geographic region as the article recommends?

It says in that deck: Distribute globally, rotate hourly, expire after 18 hours.

So it would be correct to say that if any of Cloud Flare's endpoint servers is compromised, the adversary can decrypt traffic passively on a global scale?

I assume that's the point </s>

With Anycast I don't think you have the choice of not georeplicating the STEK. And latency is one of their biggest selling points. So you could say they are trying to make the best of a bad situation.

It is absolutely bizarre that the STEK-encrypted session key is not itself sent inside the session encrypted channel.

You do have the choice. You don't need session tickets at all to run TLS. The cost of geographically-distinct STEKs would simply be an additional handshake if the client hops regions. (Is this common?)

This is a choice Cloud Flare is making in favor of performance, and it seems sort of risky with respect to a well-funded global adversary. This makes persistent access to any single endpoint server incredibly valuable.

But maybe regional STEKs are impractical from a performance perspective. I assume Cloud Flare has performance measurements to justify this choice. I'd be interested to read a blog post about it.

TL/DR: we rotate them every hour, but need to keep history of previous STEKs for 18 hours to support the maximum session lifetime:

$ openssl s_client -connect cloudflare.com:443 2>/dev/null | grep "lifetime hint" TLS session ticket lifetime hint: 64800 (seconds)

We've written about how we manage TLS session tickets here: https://blog.cloudflare.com/tls-session-resumption-full-spee....

Additionally, I wrote here about a bug that we encountered with Microsoft's implementation of TLS session resumption: https://blog.cloudflare.com/microsoft-tls-downgrade-schannel....

Here's a snippet from my blog post:

Session Tickets at CloudFlare CloudFlare’s solution to this problem, documented in previous blog posts, is to frequently regenerate and synchronize these session ticket keys across our entire global network. We currently do this once per hour. This means we need a mechanism for turning over session ticket keys. For instance, if a client instantiates an HTTPS session at 12:00pm and continues using that ticket past 1:00pm, our edge network will re-encrypt the ticket with a brand new session ticket key.

To accomplish this, our web servers must have both the full history of all previous keys that could have encrypted the ticket (i.e., one per hour dating back to the maximum session lifetime of 64,800 seconds) as well as immediate access to each newly generated key. The previous keys are used exclusively to decrypt tickets presented by the client, while the new keys are used to "refresh" the encryption on existing tickets and encrypt tickets for entirely new sessions.

These are great details, but they don't really answer my question.

One of the original goals of Google's QUIC[0] protocol was to achieve 0-RTT handshake. I wonder how TLS 1.{2,3} Session Tickets will affect the future of that protocol. Has it achieved it's purpose (applying the right pressure to the TLS working group)? Will it be phased out now?

[0] https://blog.chromium.org/2013/06/experimenting-with-quic.ht...

TLS 1.3 also supports 0 RTT. Subject to some replay attacks, the solution to which is “the application should detect and ignore that.”

I felt like Google was not going to move to TLS 1.3 if they had not included 0-RTT. Now that TLS 1.3 has agreed to include it, it seems logical to terminate QUIC crypto.

Hm, looks like maybe a ticket should be filed against:

https://github.com/ssllabs/research/wiki/SSL-and-TLS-Deploym...

As it stands, the point on resumption reads a little on the positive side?:

> 3.2 Use Session Resumption

> Session resumption is a performance-optimization technique that makes it possible to save the results of costly cryptographic operations and to reuse them for a period of time. A disabled or nonfunctional session resumption mechanism may introduce a significant performance penalty.

Does http2 have similar issues with session resumption (especially: compromised pfs)?

I came across this, which documents how cloudscape does this securely - at least they rotate - but looks like read access to memcache+compromise of a single web server is enough to compromise the past hour or two of ssl traffic that goes through cloudflare?

https://blog.cloudflare.com/tls-session-resumption-full-spee...

[ed: also came over this:

https://github.com/mozilla/cipherscan

Which looks very handy for sanity-checking servers.]

> I came across this, which documents how cloudscape does this securely - at least they rotate - but looks like read access to memcache+compromise of a single web server is enough to compromise the past hour or two of ssl traffic that goes through cloudflare?

but (depending on how you get in) you can probably also compromise the next few hours, so one extra hour doesn't seem like a huge difference, given that the scenario is somewhat unlikely in the first place.

It's relevant in terms of pfs. Say, you happen to have ciphertext of all traffic to a domain for the past three years (whitepower.forum.example.ru), and you get an urgent need to read that ciphertext. Now, if you could get physical access to any one cloudflare server, and perhaps dump the ram, or do a cold boot attack - that might have been enough to read all that data. Assume for the sake of argument, that the servers hosting the site is (physically) out of reach.

aiui getting the current keys won't magically let you decrypt those three years of traffic. that's the whole point of rotation.

Yes, I wasn't trying to say rotation is useless - just highlight that n servers still lead to n avenues of getting at traffic for all n servers, and that cloudflare did something to deal with the pfs issue. It's worse than n servers without session resumption, but better than it could be.

[ed: per https://news.ycombinator.com/item?id=15360922 the window is 18 hours]

If I understand the article correctly, are all of the issues listed addressed by TLS 1.3?

I believe you're correct.

OK, cool! In that case, I wish the article title was different, to reflect that the issues detailed in the article have all been fixed in TLS 1.3. I think it wasn't until the end of the first issue that "fixed in TLS 1.3" was mentioned.

Strongly mitigated if not entirely fixed, but TLS 1.3 deployment is stalled on awful middleboxes breaking connections they have no business in, and TLS 1.2 is not nearly broken enough to see a significant exodus.

So this is what we are stuck with for the time being.

TLS 1.3 still has problems with STEKs. If you use STEKs with 0-RTT mode, then you lose forward secrecy and that's where the most sensitive data is likely to be: your request, password, credit card number, etc ...

0-RTT doesn't have to use STEKs, there's a better way to do it, but TLS1.3 won't enforce or require it (though it could), so it'll be up to the marketplace of ideas and security standards to sort it out.

0-RTT is an awful idea anyway; the equivalent of spray-and-pray. Some folks under HTTPbis are working on an Internet Draft on Early Data's (~0-RTT's) ramifications [1] in HTTP.

0-RTT trades performance at the expense of security properties inside the same tunable protocol, which is the sort of wishy-washy stuff I (and others) were hopeful we'd get away from, the same way PFS ciphersuites went from obscure to preferred overnight, the same way cleartext HTTP has been marginalized, the same way broken ciphersuites were aggressively blacklisted and underused ciphersuites were pruned.

[1] https://tools.ietf.org/html/draft-ietf-httpbis-replay-00

I understand that TLS 1.3 is in draft stage currently. Is there an expected date of when TLS1.3 is supposed to be available in common libs and browsers ?

Also at current rate it sounds like it’s going to take years to phase out TLS 1.1. Would mordern browsers take a stand and refuse to initiate connections for older versions of TLS and its not just browsers right, there are other odd ball clients and enterprises using IE7 or something super old. I liked the way Apple had taken a stand with App Transport Security initially but even they backed down and pushed the deadline indefinitely.

We are still working around bugs in network middleware. I hope that it will be available in Chrome and on Google's servers in Q1, 2018.

Is that a euphemism for "waiting for BlueCoat shitboxes to die and get replaced"?

> Also at current rate it sounds like it’s going to take years to phase out TLS 1.1.

It's already more or less phased out. On my webservers, TLS 1.1 accounts for 0.1% of traffic, and about half of that is junk requests like exploit attempts.

Check out SSL Pulse, specifically the Protocol Support graph:

https://www.ssllabs.com/ssl-pulse/

This used to show 100% support for TLS 1.0, which is now at 92.6% as some sites are now going 1.2-only. That's just webserver support, not usage. Huge real-world difference. Like my car supports not wearing a seal-belt but I always use it. The vast majority of usage is 1.2, and a large percentage of 1.0/1.1 traffic is unwanted garbage traffic. Hence why some people are disabling 1.0/1.1 in their webservers. It also exposes more code for questionable benefit.

BTW I'd love to know what sites support TLS 1.0 but not 1.2. What's the breakdown of the Alexa Top 1000 or so? I suspect it's mostly banks and unknown sites.

> It's already more or less phased out.

Was TLS 1.1 ever "phased in"? It was the "most recent" TLS version for only a couple of years, so the early adopters went quickly to TLS 1.2, while the late adopters stayed at TLS 1.0 (or even "TLS 1.0 but disabled by default, therefore actually SSL 3.0"). Once the later adopters catch up, there's no reason for them to not jump directly to TLS 1.2.

We are doing all these session tickets gymnastic to reduce the round trip. Mostly considering web applications. In case of native mobile application, we could just add a splash-screen/quick 2 sec animation hiding the worst case 3x200ms round trip latency overhead. So at the start of the app we simply create a fresh perfectly secured TLS connection.

It appears a better choice from security standpoint, UI can always be cleverly tricked into being smooth.

It's not just reducing RTT latency; it's also reducing the number of bignum crypto operations the server needs to perform, which carry a significant computational cost.

In case of 2.5G connections (I still get 2.5G on a regular basis, e.g. in some specific train stations on my way to work in the morning) it can easily be 800ms per RTT - so 2.4 seconds before you have a chance to request any content.

And no, you cannot really ignore this if you are interested in fostering TLS' adoption in the real world. Every tenth of a second causes a measurable loss in user interest, and in many organizations this metric will drive the decision. We cannot legislate from our ivory tower and expect the world to follow against their (perceived) best interest.

For HAProxy 1.7 (may be present in older versions, didn't check), there's the 'no-ssl-reuse'[0] directive for the server.

[0]:https://cbonte.github.io/haproxy-dconv/1.7/configuration.htm...

Well, on the bright side h2 and keepalive are other ways of reducing the overall number of tls handhakes required if you're doing http.

Too bad session resumption has issues.

NewSessionTicket is sent before ChangeCipherSpec means the message is not encrypted using the master secret exchanged with the handshake and is not necessary means that the session ticket is in plaintext. Quite contrary, it has been encrypted using the server secret key. In the page 3 of RFC5077, it states "a ticket that is encrypted and integrity-protected by a key known only to the server." and in the page 11, "Tickets must be authenticated and encrypted to prevent modification or eavesdropping by an attacker."

Reply on news.ycombinator.com