Search papers, labs, and topics across Lattice.
Google DeepMind
1
0
2
A mere 1% of poisoned samples can flip classifier labels, leading to catastrophic false positives and negatives in jailbreak detection systems.
