GUIDA TECNICA

La difesa Sandwich per i prompt

The sandwich defense repeats or restates the trusted task instruction after untrusted content has been inserted into a prompt.

Ascolta questa guidaNarrato dall'IA · ~3 min

In questa pagina3 minuti di lettura
  1. Panoramica
  2. Immersione profonda
  3. Implementazione nel mondo reale
  4. The Future of The Sandwich Defense for Prompts
  5. Metterlo in pratica
  6. Continua a esplorare
  7. Domande frequenti

Panoramica

The closing reminder may reinforce the intended task, but it remains text-based guidance inside the same model context and cannot guarantee that an injection will be ignored.

Immersione profonda

The sandwich defense places untrusted content between an initial trusted task instruction and a closing repetition or restatement of that instruction. For example, a prompt may ask for a summary, insert an external page, and then remind the model to summarize the page rather than follow directions inside it. This is a prompt-template change: it does not modify the model, create a separate data channel, or restrict tool access. It is easy to test, but it should be treated as a heuristic rather than a security boundary. The intuition is that the final reminder may help the model focus on the intended task after reading the untrusted passage. That does not mean models always follow the last instruction, nor that the closing reminder structurally outranks hostile text. A malicious passage may imitate trusted instructions, contain multiple directives, or exploit behavior not covered by the template. The defense can also fail if the untrusted block enters context through another route that the prompt author did not account for. Evidence is model- and attack-dependent. A 2026 preprint tested prompt sandwiching and other prompt-level defenses against domain-camouflaged injection across three model families and three synthetic deployment domains. It found substantial variation by model, and none of the tested prompt-level defenses eliminated the threat across weaker models. The study is limited to its setup; it is evidence against assuming a universal effect, not a complete ranking for every real system. Use sandwiching only as one layer alongside explicit data handling, permission checks on tools, attack testing, and user confirmation for consequential actions. Measure both resistance and task quality before release.

Implementazione nel mondo reale

A translation prompt gives the task, inserts a user-provided paragraph, then repeats that the paragraph should be translated rather than obeyed.

A document question-answering workflow states the question, includes a retrieved passage, and restates the requested answer format before generation.

A summarizer tells the model to summarize a forum post, places the post in a marked region, then closes by reiterating that post text is source material.

A security test compares a sandwich prompt with a baseline using the same benign and malicious documents across several model versions.

The Future of The Sandwich Defense for Prompts

Prompt-level defenses will remain attractive because they are quick to prototype, but their value will depend on model behavior and the attacks a system encounters. As evaluations improve, teams may get better evidence about when repetition helps. The lasting lesson is to test this pattern within a layered design and retain deterministic limits on what an agent can do. New models may respond differently to the same closing reminder, so validate updates before relying on them. Keep safeguards outside the prompt for consequential actions.

Metterlo in pratica

Perché è importante

  • Costo e budget. Le decisioni relative all'architettura determinano prestazioni e costi operativi per anni.
  • Decisioni più chiare. La formazione tecnica aiuta i team a scegliere lo stack giusto, non solo quello più nuovo.
  • Controllo di qualità. Migliori scelte ingegneristiche riducono gli incidenti legati all’affidabilità nella produzione.

Attenzione

  • L'ottimizzazione di un benchmark può nascondere debolezze di sistema più ampie.
  • I costi delle infrastrutture e della manutenzione sono spesso sottostimati.
  • Le lacune in termini di sicurezza e osservabilità possono aumentare man mano che i sistemi diventano più complessi.

Prossimi passi

  1. Definire obiettivi di latenza, qualità e costi prima dell'implementazione.
  2. Benchmark in condizioni di carico e dati realistiche.
  3. Monitoraggio dello strumento per errori, deriva e impatto sull'utente.
  4. Preparare percorsi di rollback e risposta agli incidenti prima della scalabilità.

Continua a esplorare

Newsletter gratuita

Ricevi il briefing quotidiano sull'IA

Tre storie verificate sull'IA ogni mattina nei giorni feriali, scritte in un inglese semplice. Gratuito per sempre, senza pubblicità.

Un'email ogni giorno feriale. Disiscriviti con un clic. Non vendiamo né condividiamo mai il tuo indirizzo.

Mettiti alla prova

Take the The Sandwich Defense for Prompts quiz

Feedback immediato su ogni risposta e un certificato condivisibile con un ID verificabile una volta superato un corso.

Inizia il quiz

Sostieni l'istruzione gratuita sull'IA. AI Understanding è un'organizzazione non profit 501(c)(3) — nessun annuncio, nessun paywall, mai. Fai una donazione

Domande frequenti

What is The Sandwich Defense for Prompts?

The sandwich defense repeats or restates the trusted task instruction after untrusted content has been inserted into a prompt. The closing reminder may reinforce the intended task, but it remains text-based guidance inside the same model context and cannot guarantee that an injection will be ignored.

Where does the sandwich defense place the repeated task instruction?

The guide defines the pattern as a trusted task instruction before and after the untrusted passage.

How could a closing reminder help in this pattern?

The guide describes the reminder as a heuristic that may reinforce the task, not a guarantee.

What does sandwiching change in the application?

The guide says sandwiching is a prompt-template change, not a model or permission change.

Why is the sandwich defense not a security boundary?

The guide explains that text repetition does not create a separate channel or structural access control.

What did the cited domain-camouflage preprint evaluate?

The guide describes the preprint’s bounded setup: three model families, domains, and synthetic documents.