Skip to content
anythingai.space
AI Securityintermediate12 min · 9 steps

Data Leakage

Find the seven channels through which an AI system leaks sensitive data, and close each one with an ordinary, testable control.

IT assistantAnother useranswers, retrievalLogs & tracesraw PII keptShared cachememory across usersSystem promptinternal detailsAttacker serverinjected links / imagesModel providerretains or trainsFine-tuned weightsmemorized records
errorrequestdatatool callresponse
1 / 9
ProblemStep 1 of 9

Seven ways data walks out of an AI system

Data leakage is sensitive information reaching someone or something that shouldn't have it: another user, a log, a third party, an attacker's server. OWASP files it as LLM02 (sensitive information disclosure) and LLM07 (system prompt leakage).

Most leaks are not exotic model behaviour. They are ordinary authorization and data-handling bugs in a new pipeline: an unfiltered index, a shared cache, a debug log left on.

InsightThe most common RAG leak is an authorization bug, not a model bug: retrieval returned something the user was never allowed to read.
What changed
  • +Seven leak channels around the assistant

Lesson 10 of 22 in AI Security