LMSYS-Chat-1M

Researchers detect multi-turn LLM attacks via activation signals

Researchers have developed a new method called Latent Adversarial Detection to identify multi-turn prompt injection attacks against large language models. This technique analyzes the internal activation patterns within …