Re: [TLS] PR#625: Change alert requirements

[Ilari Liusvaara <[e-mail]>]

On Sat, Sep 10, 2016 at 08:43:33PM +1000, Martin Thomson wrote:
> 
> I wouldn't say that this is a great argument, but it's one that could
> be made.  Generally, I've given up on TLS error codes being useful, or
> even making them useful; we've been stung in the past by being overly
> specific about what went wrong.

Well, regarding usefulness of TLS error codes, the error codes don't
even come close to useful-for-debugging. This isn't because of
information leaking (which AFAICT can only happen in very few places),
but because error codes just are not expressive enough.

Also, if you have a place where you worry about alerts leaking 
information to attacker, watch out, you might very well also have a
timing attack.

The places of TLS 1.3 that I can quickly name that are sensitive,
and so one should take care to only send one kind of alert for
each:

- Deprotect failures (bad_record_mac)
- MAC (finished) failures (decrypt_error)
- key agreement failure (???, my code uses illegal_parameter[1])
- signature failures (internal_error)

Also, note that with some bad cryptographic algorithms, one can't
even let the other side know operation failed. So even just
resetting the connection leaks too much information.


[1] Logic being: It is likely provoked by peer sending invalid
share -> invalid field value -> illegal_parameter.