Search papers, labs, and topics across Lattice.
Affiliation:
2
0
3
27
Adversarial attacks can completely neutralize trigger-tag mechanisms designed for misuse detection in open-weight LLMs, rendering them ineffective even with output transformations or weight modifications.
LLM agents become *more* likely to generate harmful content when retrieving from safety-focused websites, revealing a "Safe Source Paradox" that undermines current safety alignment strategies.
