No, don't enable revocation checking

https://www.imperialviolet.org/2014/04/19/revchecking.html

by moonboots • 12 years ago

180 155 12 years ago

wpietri2• 12 years ago

For those, like me, wondering who the author might be, it appears to be this guy: "Adam Langley works on both Google’s HTTPS serving infrastructure and Google Chrome’s network stack. From the point of view of a browser, Langley has seen many HTTPS sites getting it dreadfully wrong and, from the point of view of a server, he’s part of what is probably the largest HTTPS serving system in the world - See more at: http://www.rsaconference.com/speakers/adam-langley#sthash.HM...

timothya• 12 years ago

He was also one of the two people who prepared the fix for Heartbleed in OpenSSL[0].

[0] https://github.com/openssl/openssl/commit/731f431497f463f3a2...

tptacek22• 12 years ago

He's also the author of Golang's native crypto/tls TLS stack, a longtime contributor to the IETF TLS WG, and the author of some of OpenSSL's curve software. He's not messing around.

wpietri2• 12 years ago

Let's hope he reads this and discovers the value of an "About me" sidebar on his blog.

einhverfr15• 12 years ago

On the other hand we found that as of Friday, Chrome DID NOT recognize that one of our wildcard certs for Efficito had been revoked. We sent out an email to our customers saying to enable cert revocation checking.

Revovation isn't perfect and I would not suggest the current status quo is OK but the intermediary approach Chrome takes cannot be trusted as they have now shown.

If Chome will not show our cert as revoked what is the point of revoking the cert? The author has points but the approach Google ie taking is a cure worse than the disease...

chmars• 12 years ago

Honestly, I don't see a point in certificate revocations anymore, i.e., your implicit conclusion seems to be correct. And I don't blame Google for our broken revocation system – especially because the even the best revocation system couldn't fix the current certification system that is broken in its core.

richardwhiuk2• 12 years ago

Revoking certificates is pointless - that's Adam's point here.

yuhong2• 12 years ago

I am thinking that a HSTS option enabling hard-fail OCSP plus OCSP stapling is probably a good idea, though probably less secure than putting it in the certificate.

wyager5• 12 years ago

I'm honestly kind of surprised how little action there has been to assist with a migration away from the CA model. The technology is there, but people just don't seem interested enough to leverage it.

Systems like Namecoin could serve this purpose marvelously. Powerful devices have direct access to the entire cryptographically authenticated DNS and certificate database. Weak devices can specify whom they trust to provide them with DNS/certificate data, and even those devices get some cryptographic security guarantees thanks to technologies like SPV.

dvanduzer13• 12 years ago

There's been plenty of action, but you can't turn the whole world on a dime.

Namecoin is fantastic in theory, but has the fatal flaw of using Bitcoin: the fastest number cruncher wins. Some would argue that the strength of Bitcoin's tech is that numerous currencies with different genesis blocks can flourish. That doesn't get us anywhere with naming, though.

Dead horse flog: the CA model's problem is that you can't do federated (global) naming and federated trust in the same system.

wyager5• 12 years ago

the fastest number cruncher wins.

What's the issue with that? It's reasonable to assume that the good number crunchers will always have more power than the bad number crunchers, and if that assumption ever fails, it's easy to detect and we're simply back to CA levels of security.

sarahj5• 12 years ago

Why have a single entity at all? Moxie Marlinspike proposed Convergence (https://www.youtube.com/watch?v=Z7Wl2FW2TcA) as a solution - I think that something like that has far more potential wheels to travel than a Namecoin based system.

I should be able to choose who I trust, a notary system would allow me to do just that. No central CA systems.

The biggest concern I can see is Identity management, but, as mentioned by Moxie, most of these CA don't do anything close to proper Identity management any more - I have a number of certificates bought from quite a few different CA's all made out to my rabbit, at no fixed address.

Notaries can, of course, do additional verification - they could even advertise this as a premium.

I don't see why this can't be extended to DNS lookup's either. I trust X notaries and pin the results I get, I can choose to trust a majority, or be hyper paranoid and require everyone to agree. No need to run a power hungry blockchain, no single point of technology failure.

Technically, all of that is feasible today. And I imagine we will see a number of different technologies combined to form a proper, decentralised, system.

dvanduzer13• 12 years ago

Want to help us work on this problem at http://telehash.org?

wyager5• 12 years ago

Why have a single entity at all?

How do you figure that Namecoin is a "single entity"?

vxNsr• 12 years ago

The project seems to have lost support, the last github commit was over 2 years ago.

Do you know if there was a specific reason or were people just not interested/none of the browsers jumped onboard?

SixSigma• 12 years ago

But I trust your rabbit and it can happily take as many of my dandelions as it wants.

einhverfr15• 12 years ago

I think we should rebuild a PKI on top of DNS also. It shouldn't be that hard to do honestly and it would avoid a lot of this sort of problem.

jrockway• 12 years ago

The migration away from the CA model is called "certificate pinning". Chrome uses it for high-value sites, and you use it whenever you ssh somewhere and the key's fingerprint is in your .ssh/known_hosts file.

wyager5• 12 years ago

The migration away from the CA model is called "certificate pinning".

TOFU/POP is not an effective model for the web. There are simply too many sites for it to be useful. It's pretty much an everyday occurrence that I go to a site I've never been to before, and certificate pinning won't help at all there.

rwallace• 12 years ago

If Namecoin is anywhere near as insecure as Bitcoin, it's a nonstarter. Yes, I know the cryptography underlying Bitcoin is secure, but as a matter of practical fact, Bitcoin itself as an end-user technology is hopelessly insecure. It's one thing having an endless stream of people waking up to find their bitcoins are irrevocably gone because someone hacked the computer, but we can't have domain names being irrevocably lost in the same way.

Does Namecoin actually work like that? If so, is there a similar alternative that doesn't?

akerl_• 12 years ago

The author appears to entirely ignore attack vectors where the malicious party can record but not modify/block traffic.

Edit: I get it, I missed that for sites where the key has been changed the stolen key no longer allows such eavesdropping. Thank you to yuhong for helping point this out rather than just laughing at my ignorance while pushing me down the page.

yuhong2• 12 years ago

Then a SSL MITM attack would not be possible at all.

gojomo11• 12 years ago

Still seeing lots of explanation about why the current system sucks, and not much about how a more robust system might be created and promptly adopted. Langley (the author) mentions short-lived certificates (either rapid expiration or via a 'must staple')... how soon can we enforce that? How short can that make the danger-period where the CA, and Google, and the "connected web" all know that a certificate is invalid, but a user-at-risk does not?

Why not other ways to rapid-broadcast invalidity in censorship-proof ways, so that a browser encircled by an enemy can quickly figure out something's wrong? (Or, why can't security professionals get around interdiction as effectively as copyright pirates do?)

dvanduzer13• 12 years ago

See my other post. Systems with shorter term authentication have been around for decades. The problem with X.509 is that it centralizes authorization with the browser vendor.

moe3• 12 years ago

how a more robust system might be created and promptly adopted

I'm quite fond of how the SSH host key system works.

Prompt me the first time I see a new key, provide me with supporting evidence (e.g. show me how many people have previously accepted this fingerprint for this domain) and alert me the same way in the future if the key ever changes.

If the 'supporting evidence' was plugin-based then this system could quickly become more user-friendly and trustworthy than the current centralised system can ever be.

There could be plugins to automatically trigger a SMS challenge on first contact with particularly sensitive sites. Multiple competing P2P web-of-trust plugins, plugins that let you follow trust-lists from third parties, etc.

In the current system you rely on a single, very questionable opinion on the trustworthiness of a given certificate. In the new system you'd be presented with a trust-score compiled from a whole range of opinions. The sources of which you chose before-hand.

Of course this approach doesn't include a license to print money for corrupt CA organisations and is not going to happen for that reason alone.

einhverfr15• 12 years ago

Supporting evidence, to be worthwhile, must come from a trusted source, right?

gojomo11• 12 years ago

All interesting ideas... but don't directly address rapid trust revocation, as in the case of recent relevance: a site's private keys are assumed to have been compromised (as if by the heartbleed bug).

Or are you suggesting every browser will contact many of its personal web-of-trust sources on every secure-connection? Without additional innovation, that seems just as prone to the performance bottlenecks or soft-failure (on stale data or blocked connections) as the current system.

e12e3• 12 years ago

I think the author is a little disingenuous with the term "security theatre". Basically he argues that OCSP doesn't work because hard fail might cause DOS -- but fails to conclude that without OCSP SSL/TLS is useless. It's a long argument for saying that the CA system is broken (you can only trust the white-list chrome provides) -- and the sensible conclusion is that you cannot trust any other certificate chains (without OCSP) is left out.

Without certificates, SSL/TLS falls apart.

Perhaps a better use of CAs would be to always delegate authority to the domain owner -- we'd only need OCSP for the CAs, and domain owners could issue hour/day-valid certs via a cert infrastructure. That would push a lot of complexity down to domain owners, it would probably lead to a lot of errors in implementation -- but those errors would only affect the domains -- not the main CA trust chain as such.

I'm not sure if that would be an improvement or not -- but at least you could know that if a domain was run correctly, a valid certificate could actually be trusted...

dvanduzer13• 12 years ago

The "better use of CAs" you describe is essentially DNSSEC.

tptacek22• 12 years ago

Sure, if "making the DNS totally unreliable", "baking 1990s crypto into the core of the Internet", and "conceding the CA PKI to world governments" is your idea of "better use of CAs".

e12e3• 12 years ago

No, not really. DNSSEC secures DNS, allowing it to be used (among other things) as a secure transport for delivering other certificates.

CAs already provide the same trust infrastructure, but due to incentives do not sign delegating certs by default, but typically charge extra for certs that allow the owner of a domain to set up their own internal CA.

DNSSEC only works (reasonably) for TLDs where DNSSEC is implemented, and when DNS resolvers implement checking -- many ISPs don't. Delegating CAs are already part of SSL/TLS.

DNSSEC requires fidling with the DNS information -- delegating CAs, while requiring issuing certs, only require configuration of the various services (web, imap, smtp etc) -- as per usual.

For delegating CAs to improve the security of any given domain -- some infrastructure would be needed to set up an intermediate CA. I guess for small organizations, OCSP wouldn't really be needed, as the CRLs would be small (assuming different CA for public facing services and stuff like personal certs for users). Another option would be to simply roll certificates frequently.

dvanduzer13• 12 years ago

I'm having a lot of trouble getting past: "Certificates bind a public key and an identity (commonly a DNS name) together."

X.509 certificates bind a public key and a human recognizable string (a "common name") together to create a verifiable digital identity. Over-simplified, X.509 is about solving the "I'm Spartacus" problem.

CRLs solve the "He was Spartacus" problem. I agree with the broad conclusion that CRLs aren't effective for human trust, but they are perfectly reasonable for machine trust.

Why didn't the author mention Kerberos? The default lifetime of a Kerberos ticket is designed around humans: roughly the length of a work shift in front of a computer terminal.

final edit: meta-moderation is hard

lnanek2• 12 years ago

For HTTPS to a web site, the common name is the site domain name. We're not talking about anything else here.

dvanduzer13• 12 years ago

Aha. That perhaps explains all the downvoting. But this was my objection:

Name != Identity

It's nearly impossible to have a meaningful English conversation about these problems without getting that straight.

chacham15• 12 years ago

When I hear these arguments, I always look for what is wrong with OCSP Must Staple. The author says that at the bottom it might be a solution with short lived certs, but I dont see the need for super short lived certs, only short lived OCSP staples. The author presents this as the problem:

if the attacker still has control of the site, they can hop from CA to CA getting certificates. (And they will have the full OCSP validity period to use after each revocation.)

The solution here is to not allow OCSP stapling to request a new certificate and use a full OCSP check to verify that the cert wasnt revoked.

dwightgunning2• 12 years ago

Why not hard-fail by default and give the user the option to ignore/override it? Similar to the way other certificate warnings are shown to the end-user.

x0x0• 12 years ago

answered in adam's blog; see the paragraph beginning with

   Everyone does soft-fail

dwightgunning2• 12 years ago

I guess that's true when a hard-fail causes the connection to be refused immediately by the client with no user input. In that case a DoS on the OCSP servers breaks things badly.

However what I meant to suggest is a third option. Something like hard-fail with a latch. The client should opt to fail but give the user the choice to proceed.

This would seem more desirable than the current soft-fail implementations when seem to be entirely silent to the end user.

psykovsky3• 12 years ago

That is simply not true. Firefox does hard-fail with the default settings.

pronoiac4• 12 years ago

Could we use bloom filters on the CRLs, before checking with OCSP? Maybe I should go crunch the numbers on the viability of that.

MertsA• 12 years ago

I just did and if you were okay with a 0.001 probability of false positive you could list all 500,000 (possibly way off) certificates potentially exposed through heartbleed in only 877.5KB of space. The current Chrome CRL contains 24,161 serial numbers and takes up 305.3KB of space. While it isn't a perfect fix for the revocation problem it would certainly be much better than the status quo.

One problem might be that the 0.1% of sites hit by the false positive effectively couldn't use OCSP stapling but Chrome could just first call back to Google as a CRL proxy to avoid making an OCSP request when the site stapled a valid but potentially revoked OCSP response. Then just store that response from Google for current version of CRL in the cache. End result is that the unlucky false positive sites don't have tons of unnecessary (unnecessary as far as the OCSP spec is concerned) OCSP requests going to the CAs and the only thing they would notice is that a new visitior takes 100ms longer to make the first page load.

And through the magic of bloom filters if you wanted to bump the false positive rate down to 1 in 10,000 it only bloats the list to 1.14MB. Furthermore, there are methods to make the bloom filter scale-able such that a client doesn't have to necessarily download the whole bloom filter again if a bunch of elements are added to it and instead just download a portion of the data required for a full update.

The more I think about it the more I wonder why this isn't already in Chrome in some form or another. The only downside is weird networks where OCSP might be filtered, but not https, and access to Google is filtered.

Edit: One thing I feel stupid for overlooking is that Bloom filters aren't cryptographically secure so an attacker could theoretically find a serial number for some CA that would cause a site to always be a false positive but I don't think any CAs are still giving out serial numbers in a predictable way after the MD5 debacle and even if they were it would seem to be impractical to me. The fix would just be do a SHA256 hash of the serial instead of the serial itself.

pronoiac4• 12 years ago

Oh, I wasn't thinking about having this be a perfect oracle; just a better and smaller first pass (edit:) for CRL, not for OCSP.

I got the idea from Squid and the network of caches.[1] That body of experience may be helpful.

For shrinking the size, RLE might work (most entries would be 0), and rsync may reduce bandwidth. It looks like the Squid network just used http requests for refreshes. There's probably a sweet spot for bandwidth, and I'd guess that 90-99% would work fine; you're balancing the size of the continually updated bloom filter vs. the requests for certificates that match it. I didn't worry about false positives, because it could just send an OCSP query in that case.

Your numbers for revocations sounded very low, but I just used crlset-tools[2] and checked, and it's about right. Which is weird, because someone else[3] mentioned a size of "4.107Kb" at version 1567, but that's somehow different - compression, perhaps. I thought I'd heard about CRLs megabytes long, but Google Chrome seems heavily curated re: CRLs.

I'd hash over signatures instead of the oft-predictable serial numbers, as you noted.

[1] http://wiki.squid-cache.org/SquidFaq/CacheDigests

[2] https://github.com/agl/crlset-tools

[3] https://scotthelme.co.uk/certificate-revocation-google-chrom...

pronoiac4• 12 years ago

You might check pjscott's comment, sibling to the parent comment. There's a lot more to it, but a bloom filter's already been considered for this.

pjscott• 12 years ago

Adam Langley has some thoughts about how to do this, and some numbers, here:

https://www.imperialviolet.org/2011/04/29/filters.html

pronoiac4• 12 years ago

Wow! That was an informative read. Thanks!

phunehehe0• 12 years ago

It seems the only problem with hard-fail is the risk of DoS attacks by targeting OCSP servers. However, if you include OCSP stapling you won't be affected. So a solution may be to encourage all users to enable revocation checking with hard-fail, and all servers to support OCSP stapling.

Khaine• 12 years ago

It sounds like the internet is broken Without CRL/OSCP we cannot truly trust that we are securely communicating.

Something has to give. We need to abolish SSL/TLS and migrate to something that isn't broken by design

lazyjones• 12 years ago

the internet is broken

It's not the Internet, just the CA system. There are better systems for handling trust out there, for example, people have been signing each other's PGP keys at key signing parties for decades.

einhverfr15• 12 years ago

It's not the Internet, just the CA system.

Ok, so it is just the portion of the internet that involves purchasing things with credit cards and requiring passwords to access sites. The rest of the internet is just fine.

Great. I thought for a moment that the commercial basis of the internet might be in danger. Now to determine what percentage of the internet is not dependent on the CA system.....

papaf• 12 years ago

Theres one thing that I do not understand - why not download the full revocation list?

pencilcode• 12 years ago

Potentially too much data to download, imagine, all revoked certificates in the entire internet.

sp3322• 12 years ago

You would have to crawl all OCSP providers on the internet to generate the list.

anaphor• 12 years ago

Can't the replay attack mentioned in the article be mitigated by using nonces? Why doesn't anyone do this? I'm genuinely confused by this.

colons• 12 years ago

This argument only holds if the attacker controls every internet connection you use. If you're on a portable device or you're otherwise connecting through various networks, only a subset of which are compromised, revocations are still useful.

captainmuon• 12 years ago

Exactly. If I'm on my trusted network at home and receive a big revocation list, and a few weeks later go to, say, Egypt, and someone tries to MITM me there with a stolen certificate, then it would show up as invalid.

AhtiK• 12 years ago

My quick write-up on this from few days ago, http://www.ahtik.com/blog/startssl-revocation-fees-will-not-...

Yes, revoke is broken by design, especially with mobile and Chrome browser. I'd say it's broken everywhere except Firefox with OCSP Hard Fail enabled.

Thanks to this flaw StartSSL business model has become somewhat outdated IMHO with the free certs and paid revocations.

I'm dreaming that we can fix the revocations issue with 24hour valid certificates. Suggested at the end of my post.

But I must be naive on this as it's too simple, just haven't found the flaw in this myself. Yes, it needs technical orchestration, but at least it does not add extra layer of single point of failure for every session.

EDIT: Just finished the OP post and it does indeed also mention "short-lived certificates" in the end as a potential solution.

ivanr2• 12 years ago

Indeed, short-lived certificates do seem like a solution to this problem. One downside might be the fact that (anecdotally) many users have inaccurate clocks. I read somewhere recently that a large web site has to back-date their new certificates, because, otherwise, certificate rotation/revocation causes a large spike in support tickets.

Short-lived certificates were explored in Towards Short-Lived Certificates http://crypto.stanford.edu/~dabo/pubs/abstracts/ssl-shortliv...

eps• 12 years ago

Just make sure you still check revocation of code signing certificates. Otherwise you will end up running malware that is signed with a legit key they got off my stolen Windows laptop.

mobiplayer2• 12 years ago

I've wondered many times why OCSP isn't distributed as DNS is. When we talk about websites, surely there's no more than one certificate per hostname (or less, i.e. wildcards). I don't think we're talking here of something impossible to do or not feasible with our current technology and computing power.

Also, certificate "whitelisting" could be a part of the DNS protocol itself (return the IP address of the requested hostname and the hash of its current, valid certificate).

mobiplayer2• 12 years ago

Just to clarify: OCSP is distributed, but I can't ask my local ISP OCSP server about your certificates. I have to ask your OCSP server about your certificates.

hereonbusiness2• 12 years ago

So this complete infrastructure is crap. OpenSSL, a software half the internet uses but no one cares about because it's crap. CA's not revoking keys even though they know they're compromised. Revocation being worthless because it's too much of a hassle for anyone to bother.

Great. Maybe now, when half the internet is already compromised and all our certificates are not worth the bytes they're made of ... maybe we should try to come up with something better.

edit: Actually, this whole heartbleed affair has been quite eyeopening for me, so I'm thankful for that. But it certainly didn't help with the paranoia I feel the last couple of years while using services on the internet.

tptacek22• 12 years ago

Yes! Now it's time for us to generate a whole new broken infrastructure! I'm sure if we just rewrite all the Internet's crypto in Rust, everything will be great 10 years from now. No way will a radically different new transport cryptosystem grant researchers 100 new bugs to play with; after all, we'll have option types.

hereonbusiness2• 12 years ago

I'm no security expert but I guess there could be ways to keep TLS as a protocol more or less unchanged while fixing the obviously broken stuff surrounding it.

lucian1900• 12 years ago

You're right to mock the attitude people have that the only thing wrong with OpenSSL is the language it's written in, but memory unsafety has nevertheless been a factor in many security flaws.

cpeterso• 12 years ago

Don't forget that 90% of the world's certificates are issued by five commercial CAs, who happen to be friendly with various national security agencies.

Splendor4• 12 years ago

Let me get this straight. Sites across the internet are (hopefully) revoking their CAs and issuing new ones to address Heartbleed but Mr. Langley is suggesting that we shouldn't check for revoked CAs because it might not do anything and it's slow?

Sorry, but after the last few weeks I'll happily accept a little slowness for the security revocation checking provides in the cases where it does work, even if it's not 100% of the cases.

ars2• 12 years ago

I guess you didn't read the article? He's saying there are no cases where it works. Making it completely pointless.

millerc• 12 years ago

That's why I claim that online revocation checking is useless - because it doesn't stop attacks.

Doesn't mean there are "no" cases where it works. It just means any attacker dedicated enough can work around the CRLs.

I don't see any reason why one should throw the baby with the water. In this case, I just see Chrome guilty of FUD and hiding behind an intractable problem to justify their incorrect position.

See http://en.wikipedia.org/wiki/Two_Generals'_Problem if you want to convince yourself the problem is intractable.

Splendor4• 12 years ago

I did read the article.

"In order to end on a positive note, I'll mention a case where online revocation checking does work..."

takeda• 12 years ago

Well, his argument is also that the attacker can easily circumvent it, which is true, but it is still makes it slightly harder to do, because the attacker needs to remember it.

mpyne3• 12 years ago

Well, this is what "security theater" is. If you said that exact thing in the context of a TSA screening program there would be no one here going "yes, that makes perfect sense", and it's even easier for network attackers; they have to fix their attack scripts once and they work for good until the next countermeasure.

fragmede• 12 years ago

The article gives two reasons for why 'soft-fail' is required: Captive-portals, and OCSP server failure.

To deal with captive portals: have an SSL signed 'subdomain.google.com/you_are_on_the_internet' site/page that Google Chrome can use to check to see if it's captive or not. If it's captive, enable soft-fail. If internet access is available, set to hard-fail.

Websites these days are complex, with many (digital) moving parts - the database server(s), the static image server(s), dynamic response server(s), gateway server, probably a memcache server or something similar. If any one of those goes down, the site is unusable. Why then, should the OCSP server going down be considered any differently? Is a black-hat rented bot-net running a DDoS going to care if it's the main gateway server or the OCSP server?

But let's say we do consider disabled OCSP servers to be a client-side issue. Google could query and cache the OCSP server status, either with OCSP stapling or via some side-channel they build into Google Chrome.

The combination of both would allow hard-fail to be an option in Google Chrome.

rdl• 12 years ago

Why are we not using OCSP Must Staple right now?

No, don't enable revocation checking

Related Stories