Embedding-based defenses fail against crafted attacks in LLM MAS; confidence scores from logits improve robustness but decay over communication rounds.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CR 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
When Embedding-Based Defenses Fail: Rethinking Safety in LLM-Based Multi-Agent Systems
Embedding-based defenses fail against crafted attacks in LLM MAS; confidence scores from logits improve robustness but decay over communication rounds.