Search papers, labs, and topics across Lattice.
1
2
Turns out, state-of-the-art prompt injection defenses aren't as robust as we thought: they crumble against adaptive attacks and struggle when the injected task aligns with the intended one.
