Search papers, labs, and topics across Lattice.
This paper demonstrates a novel attack against Introspection Adapters, a defense mechanism designed to improve the interpretability and safety of large language models. The attack leverages symmetries in the adapter's architecture to craft adversarial inputs that bypass the intended introspection capabilities. By exploiting these symmetries, the attack effectively blinds the adapter, rendering it unable to detect or mitigate harmful outputs.
Introspection Adapters, a promising approach to LLM safety, can be completely defeated by exploiting architectural symmetries.
We demonstrate an attack on Introspection Adapters (Shenoy et al., 2026).
BAIR