But what attack is that blunting?
1. A bad actor with access to the system will just regenerate another LE cert and carry on.
2. A bad actor who has injected a bad cert, but without system access, still has three months to conduct his fraud which is more than enough time to phish 90% of the user-base.
I'm not privy to the LE decision-making process but 90 days seems like the result of a bad compromise. Automated certs should be renewed on a very-high-frequency ( like hourly ) otherwise there's little benefit over the traditional model, and a lot of downsides such as breaking pinning.
90 days seems to have been chosen to be painful enough to push people into adopting auto-renewal but not frequent enough to overload LE's system, and thus not actually providing any security gains.
FYI: You can renew your cert without changing the key, e.g. for mail servers and DANE/TLSA.
http://www.internetsociety.org/deploy360/blog/2016/03/lets-e...
http://www.internetsociety.org/deploy360/blog/2016/01/lets-e... (part 1)
Broadly speaking Certificate Revocation is garbage. It doesn't work. So when a bad actor generates a bad certificate there's a legitimate chance that without Certificate Pinning, that Certificate will be accepted even if theoretically revoked. A three months max duration limits the potential liabilities to three months.
Additionally three months offers the better theoretical security against state actors. While a 2048 bit key is normally considered to take tens of years to break, it is possible a state actor with unlimited resources could construct a quantum computer that could cut that time down significantly. If the certificate pair is regenerated every three months using a safe cryptographic source this makes breaking it likely not worth while even for a quantum computer.