Found 1 document(s)
Page 1 of 1
-
Large language models have been reported to encode the harmfulness of a request separately from their decision to refuse it, and latent safety monitors such as Latent Guard build on this report by classifying prompts as ...