Algorithm agility is bad. You see it happening at the beginning of every TLS session; what you don't see are the long trail of vulnerabilities that have resulted from TLS's overly complicated handshake, and, worse, the essentially nonexistent vulnerabilities algorithm agility has mitigated over TLS's lifetime; in general, when TLS constructions have failed, they've failed in the joinery (block cipher modes, MAC arrangements, etc), and, most importantly, have required updates to the entire protocol anyways.
It's a bad idea and people shouldn't build it into new systems. Version the protocol instead, and plan on making it straightforward to upgrade the protocol and lock out old versions.
And all of this is to say nothing of the frankly embarrassing problems that have plagued JWT as a result of algorithm agility (alg=none). Removing agility from JWT wouldn't make it a good specification, but it would certainly make it a better specification.
This is a really tough question to answer, because the answer depends on what you're using JWT for. JWT crams as much functionality into the format as possible, and most of that functionality isn't needed for most use cases. This means that offering an alternative requires knowing some context about what you need out of JWT in the first place.
That being said, for most purposes, you can do worse than using either mutual TLS or Macaroons [0]. As always with cryptography though, the devil is in the details, so for a more thorough discussion, check out @tptacek's "A Child's Garden of Inter-Service Authentication Schemes" [1]. It's one of my favourite treatments of the topic, and discusses the tradeoffs of a few different techniques for different use-cases.
I disagree with locking out old versions. It's a huge problem with older devices trying to access the web via SSL. I think it is important to consider device longevity and also the possibility of the user not being able to upgrade their software. Security is important, but it is not everything. There is also accessibility and compatibility to consider.
It's why I still test my websites with Netscape 3.x and keep http running.
> Be conservative in what you send, be liberal in what you accept.
The first half makes good sense. You should make efforts to ensure your program is producing well-formed output. That's just a statement of software quality.
The second half is dependent on the problem domain. A web browser should do its best to cope with malformed input, sure. An Ada compiler should not, as an important part of its job is to reject erroneous inputs. Similarly, part of the value of XML schemas is to (at least partially) automate the process of detecting invalid data, perhaps so that bad requests can be rejected.
Kerberos/AD is another example of this kind of mess. It has both LDAP and LDAPS not to mention the zoo of encryption types available for keytabs and credential caches. You can still use unsalted encryption types and DES!
While on one hand insecure LDAP is convenient for testing, I do think it should really just be removed and require LDAPS.
If someone suddenly breaks the stuff based on the 25519 curve there are going to be a lot of very unhappy people. I am still waiting for an explanation of why this particular curve is immune to such a breakthrough.
Curve25519 has only been around for something like 10 years. A catastrophic attack on it would presumably involve new ideas specific to that particular or class of curves.
Otherwise any completely new system would have to be considered secure, as it would not suffer from any of the issues of any of the older systems. My understanding is that new elliptic curves can have entirely new characteristics not shared by older curves.
Split the difference with base 32? More efficient than hex, you can leave out the 4 easiest letters to confuse with digits and no upper/lower case problem.
Verbal communication is a problem with any system that uses letters. In hex for example; 3, B, C, D and E all sound very much the same.
I'm not a any kind of crypto expert, but I'm wondering if this looks like a config file? Let's say you have a JSON file that has one field, which is the key, base64-encoded.
Maybe you won't change the algorithm, but still want to leave room to add another config parameter. The file format is obvious in a text editor and it's easy to add more stuff.
21 comments
[ 197 ms ] story [ 97.9 ms ] threadIt's a bad idea and people shouldn't build it into new systems. Version the protocol instead, and plan on making it straightforward to upgrade the protocol and lock out old versions.
That being said, for most purposes, you can do worse than using either mutual TLS or Macaroons [0]. As always with cryptography though, the devil is in the details, so for a more thorough discussion, check out @tptacek's "A Child's Garden of Inter-Service Authentication Schemes" [1]. It's one of my favourite treatments of the topic, and discusses the tradeoffs of a few different techniques for different use-cases.
[0] https://en.wikipedia.org/wiki/Macaroons_(computer_science)
[1] https://latacora.micro.blog/a-childs-garden/
It's why I still test my websites with Netscape 3.x and keep http running.
> Be conservative in what you send, be liberal in what you accept.
The first half makes good sense. You should make efforts to ensure your program is producing well-formed output. That's just a statement of software quality.
The second half is dependent on the problem domain. A web browser should do its best to cope with malformed input, sure. An Ada compiler should not, as an important part of its job is to reject erroneous inputs. Similarly, part of the value of XML schemas is to (at least partially) automate the process of detecting invalid data, perhaps so that bad requests can be rejected.
edit Turns out others have spent more time thinking about this than I have: https://en.wikipedia.org/wiki/Robustness_principle#Criticism
I don’t think it’s just a crypto thing, agility is an issue in protocols that need to remain compatible in general.
While on one hand insecure LDAP is convenient for testing, I do think it should really just be removed and require LDAPS.
See https://safecurves.cr.yp.to/
Otherwise any completely new system would have to be considered secure, as it would not suffer from any of the issues of any of the older systems. My understanding is that new elliptic curves can have entirely new characteristics not shared by older curves.
So Curve25519 has had 16 years of widespread academic scrutiny. That’s “battle-tested” by modern cryptographic standards.
I would argue given the key lengths of ed25519, not even to do base64, but just do hex encoding.
Sure it is a little longer, but in return you get much simpler encoding and decoding, no worries about flavors of base64 or padding.
In addition, hex encoding is much easier to verbally communicate to someone versus base64.
In a security context, I think making things as simple as possible has big benefits.
Verbal communication is a problem with any system that uses letters. In hex for example; 3, B, C, D and E all sound very much the same.
We have a standard way of dealing with that in the form the NATO phonetic alphabet.
So 3BCDE would be “Tree Bravo Charlie Delta Echo”
Maybe you won't change the algorithm, but still want to leave room to add another config parameter. The file format is obvious in a text editor and it's easy to add more stuff.