Search papers, labs, and topics across Lattice.
1
0
2
13
LLM agents can be tricked into ignoring user instructions and misusing tools in over 90% of trials via a new "Memory Control Flow Attack" that exploits persistent memory influence.