According to the post, for the Claude language model the multilayer protection — model training, input probing and an intent-checking classifier — reduces indirect prompt injection from unknown attacks to about zero; the automatic mode will become the default in Claude's code starting next week.
AI-generated text
With multilayer protection, indirect prompt injection against Claude can be reduced to virtually zero
According to the post, for the Claude language model the multilayer protection — model training, input probing and an intent-checking classifier — reduces indirect prompt injection from unknown…
