I'm not sure how feasible this is, particularly short term, but a suggestion:
Every compiled script generates bytecode when compiled, right? And an exact copy of this script should generate an identical bytecode, I assume. Why not produce a fingerprint of these bytecodes at compile time as well, as an attack countermeasure.
When an attack is detected, send a signal to all simulators on the grid to blacklist scripts with the Fingerprint of the offending script. This will stop the attack. Then the normal procedures for handling attack objects can happen at the Asset Server level as load allows.
-Kris