Search papers, labs, and topics across Lattice.
Microsoft
1
0
2
3
Encoded adversarial prompts can exploit AI safety mechanisms with alarming effectiveness, revealing significant vulnerabilities in current models.
