Do any of the open weight models from smaller labs exist if they can't distill from the SoTA models that are throwing billions of dollars of compute into pretraining?

Well - I can't say they have always said this - but at least for Circa 1998 CCNP onwards that's been their position. The instructors were very adamant - to the point that I'm recalling this 27+ years later.

This is the first thing that as a Network Engineer I was taught - and every formal security class I've taken (typically from Cisco - they have awesome course) - repeats the same thing.

I believe the common knowledge is somewhat more nuanced than people would have you believe

I present to you two separate high-value targets whose IP address has leaked:

IPv4 Target: 192.168.0.1
IPv6 Target: 2001:1868:209:FFFD:0013:50FF:FE12:3456

Target #1 has an additional level of security in that you need to figure out how to route to that IP address, and heck - who it even belongs to.

Target #2 gives aways 90% of the game at attacking it (we even leak some device specific information, so you know precisely where it's weak points are)

Also - while IPv6 lacks NAT, it certainly has a very effective Prefix-translation mechanism which is the best of both worlds:

Here is a real world target:

FDC2:1045:3216:0001:0013:50FF:FE12:3456

You are going to have a tough time routing to it - but it can transparently access anything on the internet - either natively or through a Prefix-translation target should you wish to go that direction.