Search papers, labs, and topics across Lattice.
University of Luebeck
4
0
5
FloatDoor reveals that LLMs can be covertly compromised to exhibit malicious behavior on specific platforms while appearing benign elsewhere, exposing a critical security gap in AI deployment.
Ignoring prompt knowledge is a critical security flaw, as LLMs can covertly transmit hidden messages through their deterministic sampling processes.
Refusals from LLMs can be transformed into supportive communications that not only prevent harm but also guide users toward helpful resources.
Turns out, even hardware-protected enclaves can't stop a clever side-channel attack from stealing your decision tree models.
