Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
1
0
1
3
A staggering 77% of confident classifier decisions are susceptible to manipulation by a single token, revealing critical vulnerabilities in prompt injection detection.