Search papers, labs, and topics across Lattice.
1
0
2
LLM agents are shockingly susceptible to memory manipulation, with over 90% of trials vulnerable to attacks that force unintended tool usage and persistent behavioral deviations.
