Search papers, labs, and topics across Lattice.
1
0
2
It is shown that steering creates a predictable prompt-injection attack surface whose vulnerability increases monotonically with steering magnitude, relevant to agent systems deploying steered classifiers.
