I think I may have erred by trying to write a treatise on which algorithms are vulnerable :) Here's some updated text, trying to be more concise.
Jim: Your points about SHA-256 vs. SHA-512/256 and SHA-256 vs. SHA-3 don't really apply, since JOSE hasn't defined algorithm identifiers for SHA-512/256 or SHA-3.
"""
# Signature Algorithm Protection
In
some usages of JWS, there is a risk of algorithm substitution attacks,
in which an attacker can use an existing signature value with a
different signature algorithm to make it appear that a signer has signed
something that he actually has not. These attacks have been discussed
in detail in the context of CMS {{RFC 6211}}. The risk arises when all
of the following are true:
* Verifiers of a signature support multiple algorithms of different strengths
*
Given an existing signature, an attacker can find another payload that
produces the same signature value with a weaker algorithm
* In particular, the payload crafted by the attacker is valid in a given application-layer context
For
example, suppose a verifier is willing to accept both "PS256" and "PS384"
as "alg" values, and a signer creates a signature using "PS256". If
the attacker can craft a payload that results in the same signature with SHA-256 as
the signature with SHA-384 of the legitimate payload, then the "PS256" signature
over the bogus payload will be the same as the "PS384" signature over
the legitimate payload.
There are several ways for an application using JOSE to mitigate algorithm substitution attacks
The simplest mitigation is to not accept signatures using vulnerable algorithms: Algorithm
substitution attacks do not arise for all signature algorithms. The only algorithms defined in JWA {{I-D.ietf-jose-json-web-algorithms}}
that may be vulnerable to algorithm substitution attacks is RSA-PSS ("PS256", etc.). An implementation that does not support RSA-PSS is not
vulnerable to algorithm substitution attacks. (Obviously, if other algorithms are added, then they may introduce new risks.)
In addition, substitution attacks are only feasible if an attacker can compute pre-images for the weakest hash function accepted by the recipient. All JOSE algorithms use SHA-2 hashes, for which there are no known pre-image attacks as of this writing. Until there begin to be attacks against SHA-2 hashes, even a JOSE implementation that supports PSS is safe from substitution attacks.
Without restricting algorithms, there are also mitigations at the JOSE and application layer: At the level of JOSE, an application could require that the "alg" parameter be carried in the protected header. (This is the approach taken by RFC 6211.) The application could also include a field reflecting the algorithm in the application payload,
and require that it be matched with the "alg" parameter during
verification. (This is the approach taken by PKIX {{RFC5280}}.)
Of
these mitigations, the only sure solution is the first, not to accept vulnerable algorithms. Signing over
the "alg" parameter (directly or indirectly) only makes the attacker's
work more difficult, by requiring that the bogus payload also contain
bogus information about the signing algorithm. They do not prevent
attack by a sufficiently powerful attacker.
"""