Search papers, labs, and topics across Lattice.
3
0
7
LLMs have dramatically different and surprisingly structured safety landscapes, with some models exhibiting near-universal vulnerability plateaus while others show fragmented basins of failure.
Hubness poisoning, a subtle but critical vulnerability in RAG systems, can now be effectively detected with HubScan, an open-source scanner that spots adversarial hubs with high recall and precision.
Code Execution MCPs slash token usage and latency in agent systems, but open a Pandora's Box of new attack vectors, demanding layered defenses.
