SIGNAL//DESK
AI securitysrc: MITRE ATLAS

LLM Data Leakage

LLM Data Leakage is when a chatbot accidentally reveals secret information it shouldn't, like private messages or company files, because someone tricked it into sharing things it was supposed to keep hidden.

LLM Data Leakage occurs when an adversary uses prompt injection or adversarial inputs to bypass safety filters, causing the model to output sensitive data from its training set, connected databases, or other users' sessions.

LLM Data Leakage is a security vulnerability where adversarial prompt engineering induces a model to violate confidentiality constraints, resulting in the unauthorized exfiltration of sensitive information. This includes the extraction of memorized proprietary training data, unauthorized access to context-augmented data sources, or the cross-tenant leakage of PII and proprietary information from shared memory buffers.


← all terms