Skip to content
Vol. I · No. 251
Mon · 8 Jun
A Daily Lexicon of Trustworthy Data
The Lexicon

006·341

prompt injection

/prɒmpt ɪnˈdʒɛkʃən/ - n.

1 [colloq.] Discovering the assistant trusts the document more than the policy that approved it.Keep. Punchy.This is the problem.

Working definition

2. An attack that smuggles adversarial instructions into a model's input so it overrides its intended task.

Evidence
See also
  • agentic systemsSoftware empowered to act on decisions no human had been assigned to make.
  • guardrailsRules that enforce a policy, written for the policy the org has not yet enforced anywhere else.
  • prompt engineeringWriting the requirements doc you skipped, one sentence at a time, into a text box.
  • retrieval-augmented generationA pipeline that retrieves the company's contradictory documents and asks the model to sound certain about all of them.