[openssl-users] Personal CA: are cert serial numbers critical?

[Erwann Abalea]

Bonjour,

> Le 16 août 2017 à 16:51, Jakob Bohm <jb-openssl at wisemo.com> a écrit :
> 
> On 16/08/2017 16:32, Tom Browder wrote:
>> On Wed, Aug 16, 2017 at 08:36 Salz, Rich via openssl-users <openssl-users at openssl.org <mailto:openssl-users at openssl.org>> wrote:
>> 
>>    ➢ So, in summary, do I need to ensure cert serial numbers are
>>    unique for my CA?
>> 
>>    Why would you not?  The specifications require it, but those
>>    specifications are for interoperability. If nobody is ever going
>>    to see your certs, then who cares what’s in them?
>> 
>> 
>> Well, I do like to abide by specs, and they will be used in various browsers, so I think I will continue the unique serial numbering.
>> 
>> Thanks, Rich.
> 
> Modern browsers increasingly presume that such private CAs behave exactly
> like the public CAs regulated through the CA/Browsers Forum (CAB/F) and
> the per-browser root CA inclusion programs (the administrative processes
> that determine which CAs are listed in browsers by default).
> 
> Among the relevant requirements now needed:
> 
> - Serial numbers are *exactly* 20 bytes (153 to 159 bits) both as standalone
> numbers and as DER-encoded numbers.  Note that this is not the default in
> the openssl ca program.

There’s no such requirement. It MUST be at most 20 octets long.

> - Serial numbers contain cryptographically strong random bits, currently at
> least 64 random bits, though it is best if the entire serial number looks
> random from the outside.  This is not implemented by the openssl ca program.

It can be easily done by an external script. Generate a secret key for your CA for a 128bits block cipher, keep the monotonic counter, but instead of putting the value of the counter in the serial number, encrypt the counter with the key+cipher, and use the resulting cipher text as the serial number. Of course, increment the counter.

> - Certificates are valid for at most 2 years (actually 825 days).
> 
> - SHA-1 (and other weak algorithms such as MD5) are no longer permitted and
> is already disappearing from Browser code.
> 
> - RSA shorter than 2048 bits (and other weak settings such as equally short
> DSA keys) are no longer permitted and is already disappearing from Browser
> code.
> 
> - If the certificate is issued to an e-mail address, that e-mail address must
> also be listed as an rfc822Name in a "Subject Alternative Name" certificate
> extension.
> 
> - End user certificates must be issued from an Intermediary CA whose
> certificate is is in turn issued from a longer term root CA.

I doubt browsers impose this on purely private CAs.

> - If revocation is implemented (it should be, just in case someone gets their
> computer or other key storage hacked/stolen), it needs to support OCSP, but
> should ideally do so without having the actual CA keys online (standard trick:
> Issue 3-month non-revocable OCSP-signing certificates and provide the
> corresponding private key to the server running the OCSP responder program).
>  I would recommend to also implement traditional CRLs, since for smaller CAs
> it is a better solution for browsers and servers that support it.

Another requirement is that a TLS server certificate shall have its identity (FQDN) in the SAN extension. Use of the commonName attribute has been deprecated long ago.

Cordialement,
Erwann Abalea