ChannelGuard stops all attacks in tool output in its tests

The defense uses no more model calls and no training, but attacks that use other text can go around it.

Claimed, not confirmed