← AI security in plain English

AML.T0082

MITRE ATLAS / RAG Credential Harvesting

Adversaries use their access to a language model on the victim's system to collect credentials, prompting it to surface secrets that were indexed into its retrieval corpus.

Think of it likeAsking the new filing clerk to read out everything in the drawer marked confidential, because nobody told them not to.

In plain English

If secrets ended up in the documents a RAG system indexed, the model will hand them over when asked the right way. The indexing step is where the exposure happened.