Banana Navy
Catalog FR
Lab · Threat modeling IA · Fiche détaillée

Data and context leakage

the attacker draws data from another user or from the system's memory (data leakage, system prompt extraction, exfiltration) without necessarily cracking anything.

SheetF9
CategoryI-TB7-01 (information disclosure)
Layers14 · LLM Interpretation, 18 · Audit Logging, 15 · Policy Engine
Systemvoicebot IA

isolation of calls and of the knowledge base (RAG), least access, never cross-session access; per-requester source segregation; output redaction.

The threat

the attacker draws data from another user or from the system's memory (data leakage, system prompt extraction, exfiltration) without necessarily cracking anything.

Angle mortWhy classic frameworks miss it

the boundary is no longer network or storage but session and reasoning: the system's memory (RAG, history, context) can leak between users or be extracted by a question.

MitigationProposed approach

isolation of calls and of the knowledge base (RAG), least access, never cross-session access; per-requester source segregation; output redaction.

The proposed control
no cross-session access and no system prompt extraction.

Expected evidence
prove that one session cannot read another's memory or extract the system prompt.

SourcesReferences and public research

OWASP GenAIOWASP GenAI LLM02:2025 Sensitive Information Disclosure
Recherche publiquePublic research sources: MITRE ATLAS 2026.07 (verified technique mapping), OWASP GenAI (OWASP), and the public risk-voicebot (aivansoul/risk-voicebot) template defining the 20 checkpoints. No client registry data: generic sheet, no rating, no verdict.
couche 14 · LLM Interpretationcouche 18 · Audit Loggingcouche 15 · Policy Engine

Explore the 20 security layers

MITRE ATLAS 2026.07 · OWASP GenAI · risk-voicebot