Search papers, labs, and topics across Lattice.
Independent Researcher
2
0
5
3
Explicitly invoking external image tools in vision-language models dramatically reduces jailbreak success rates, even when the tool's output is overridden or unsafe.
LLM agents can be tricked into unauthorized actions via indirect prompt injection, but AuthGraph's dual-graph approach slashes attack success rates from 40% to 1% while preserving utility.
