Goldowsky-Dill et al.

A BASE fellowship project called SPEC-GAP found that linear probes can detect adversarial shifts in multi-agent language models before unsafe actions become apparent in outputs, but the signal is thin…