Search papers, labs, and topics across Lattice.
5
0
6
0
Prior interactions can boost LLM compliance on cybersecurity requests by over 23%, but dialogue decomposition drastically undermines this effect.
Multi-agent LLM systems are more vulnerable than previously thought, with systemic failures that local checks can't catch, demanding a new framework for security analysis.
Graph-augmented LLMs fail to leverage provided graph evidence effectively, revealing a critical gap in their design that can be addressed with targeted interventions.
Proton reveals 23 zero-day vulnerabilities in Electron apps, with 22 enabling OS command execution, exposing critical security flaws in popular software frameworks.
Mystra achieves 95.5% recall on dynamic taint analysis with zero false positives while significantly reducing runtime overhead compared to existing tools.
