Search papers, labs, and topics across Lattice.
Northeastern University, China 4 University of Chinese Academy of Sciences, China
3
0
5
Malicious instructions hidden in images can bypass existing skill scanners, exposing a critical vulnerability in LLM-based systems.
Classical Chinese, with its conciseness and obscurity, unlocks a surprisingly effective attack vector against LLM safety filters, and can be automatically exploited via bio-inspired optimization.
Coding agents are vulnerable to a new class of stealthy, automated prompt injection attacks via poisoned skills, achieving high success rates even in realistic software engineering tasks.
