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

Direct prompt injection

a user forces the model to ignore its instructions by asking directly in the conversation (ignore previous instructions).

SheetF15
CategoryAI1-TB7 (prompt injection directe)
Layers10 · Prompt Injection Filter, 14 · LLM Interpretation
Systemvoicebot IA

compartmentalize instructions (system or user separation) and above all test whether the injection actually reaches a critical workflow: if the toxic output is blocked downstream, the threat is contained.

The threat

a user forces the model to ignore its instructions by asking directly in the conversation (ignore previous instructions).

Angle mortWhy classic frameworks miss it

frameworks do not model a conversation input whose content has authority over the processing engine's behavior.

MitigationProposed approach

compartmentalize instructions (system or user separation) and above all test whether the injection actually reaches a critical workflow: if the toxic output is blocked downstream, the threat is contained.

The proposed control
no unfiltered output reaches a critical workflow.

Expected evidence
the trajectory of an injection up to an effect on the decision path (or a block).

SourcesReferences and public research

MITRE ATLAS 2026.07AML.T0051.000 Direct
OWASP GenAIOWASP GenAI LLM01:2025 Prompt Injection
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 10 · Prompt Injection Filtercouche 14 · LLM Interpretation

Explore the 20 security layers

MITRE ATLAS 2026.07 · OWASP GenAI · risk-voicebot