Traditional parsers have explicit data types; language models infer intent from text.
PROMPT & CONTEXTAI SECURITY ENGINEERING
Instruction/data confusion
LLMs process instructions and data in the same context, creating ambiguity about what should be obeyed.
Tool output that says 'ignore previous instructions' can influence the next reasoning step unless treated as untrusted.
Which security control should be explicit?
REMEMBERText provenance matters even when the model cannot inherently enforce it.
No uploads · No company data · No account required