[akerl_]

How would that work in practice? If I compromise a pile of IoT devices running on Comcast users' networks, and use them to launch and attack, all Comcast users on their subnet get marked as uncool? And if we're marking them as "bad", doesn't that mean all of their BGP peers mark them as uncool and then the weights for their prefix are lower but still even, so routing still ends up the same?

The only way they'd be impacted would be if some networks didn't implement your bad-actor-prefix-weight-mod, and then we'd just be penalizing the people who don't use your system along with the attackers, since we'd be routing the bad traffic via their networks.

[voltagex_]

You can see the impact of this kind of thinking in RBLs and blocklists - try to send email via your residential connection and you probably won't be able to.