Search papers, labs, and topics across Lattice.
1
0
2
LLM agents are surprisingly vulnerable to "silent egress" attacks, where malicious websites trick them into leaking sensitive data through hidden outbound requests, even when the final output appears safe.
