What Is Prompt Injection?
Prompt injection is an attack where input text tricks an AI model into ignoring its instructions. Learn direct vs indirect injection and voice agent defences.
Réponse rapide
Prompt injection is an attack in which input text makes an AI model behave in ways its owner did not intend. OWASP defines it as user prompts altering a large language model's behavior or output in unintended ways. It can be direct, from the user, or indirect, hidden in content the model reads. A key defence is limiting what the agent can do.
An AI voice agent follows instructions written by its owner, then listens to a stranger. Prompt injection is what happens when the stranger's words compete with the owner's instructions. This page explains the term and the sensible, non-technical ways to reduce the risk.
What Is Prompt Injection?
OWASP's Top 10 for LLM Applications 2025 lists prompt injection as LLM01. It says a prompt injection vulnerability occurs when user prompts alter the LLM's behavior or output in unintended ways. NIST's Generative AI Profile (NIST AI 600-1) describes it as modifying what input is provided to a generative AI system so that it behaves in unintended ways.
The root cause is that the model reads instructions and data in the same stream of text. It cannot always tell which words are the rules and which are the request.
Direct vs Indirect Prompt Injection
| Direct | Indirect | |
|---|---|---|
| Where it comes from | The user's own input | External content the model reads, such as websites or files |
| Who is involved | The person talking to the model | A third party who placed text where the model will retrieve it |
| Voice agent example | A caller says something meant to override the agent's role | A document or page the agent retrieves contains misleading instructions |
| Main defence | Constrain the role and limit what the agent can do | Control and mark what the agent can read |
NIST describes indirect injection as adversaries exploiting LLM-integrated applications by injecting prompts into data likely to be retrieved. OWASP notes that, like direct injection, indirect injection can be intentional or unintentional.
How Could a Caller Try It on a Voice Agent?
A caller can only speak, so the attempts are ordinary sentences, for example telling the agent to forget its instructions, asking it to read out its own rules, claiming to be a manager and demanding a discount, or asking for something outside its job. These examples are deliberately general. The point is not the wording but that anything a caller says becomes input to the model.
The related risk is a confident wrong answer without any attacker, covered under AI hallucination.
How Do You Reduce the Risk?
- Give the agent a narrow job. The first measure OWASP lists is to constrain model behavior by defining its role and limits.
- Limit what it can do. OWASP recommends least-privilege access: the fewer actions the agent can take, the less a trick can achieve.
- Scope the knowledge base. Answers should come from your own documents, as described in the knowledge base guide.
- Keep a human route. OWASP recommends human approval for high-risk actions. On a call that means handing off to a person; see human hand-off.
- Test and review real calls. OWASP recommends adversarial testing. Follow how to QA AI voice agent calls and listen to recordings after launch.
- Know where the data goes. Lisez le politique de données IA et le page de confiance.
DialerBee agent vocal IA speaks only prices in your catalogue, answers from your uploaded documents and can be stopped at the scope you choose with kill switches. These are controls that reduce risk. No control makes a language model immune: OWASP says it is unclear whether fool-proof prevention methods exist.
Foire aux questions
What is prompt injection in simple terms?
It is input that changes what an AI model does in a way its owner did not intend. OWASP defines it as a vulnerability that occurs when user prompts alter the LLM's behavior or output in unintended ways.
What is the difference between direct and indirect prompt injection?
Direct injection comes from the user's own input. Indirect injection comes from external content the model reads, such as websites or files.
Can a caller prompt-inject a voice agent?
A caller's spoken words become text the model reads, so a caller can try. The practical defence is to limit what the agent can do, so that a successful trick has little to reach.
Can prompt injection be fully prevented?
OWASP notes that, because of the stochastic nature of generative models, it is unclear whether fully reliable prevention exists. The measures aim to reduce the impact.
Does a knowledge base make a voice agent safer?
A scoped knowledge base keeps answers tied to your documents, which helps. Treat any external content the agent reads as untrusted, and review real calls.
How do I test for it?
OWASP lists adversarial testing and attack simulations among its mitigations. For a voice agent, that means placing test calls that try to push the agent off its role, and listening to what it does.
Sources
- OWASP Top 10 for LLM Applications 2025, LLM01:2025 Prompt Injection (source, checked 2026-11-19)
- NIST AI 600-1, Generative AI Profile, July 2024 (source, checked 2026-11-19)
À lire aussi : AI hallucination · AI voice agent human hand-off · AI voice agent knowledge base · QA of AI voice agent calls · Confiance
This page is general information, not legal advice. DialerBee does not provide legal advice or guarantee regulatory compliance.
À lire aussi : What Is Automatic Speech Recognition (ASR)? · What Is a Large Language Model (LLM)?
Termes associés
What Is Regulation F? The 7-in-7 Call Rule
Regulation F is the CFPB rule that implements the FDCPA for debt collectors. Learn the 7-in-7 call frequency presumption and how it shapes collection dialing.
What Is Retrieval-Augmented Generation (RAG)?
Retrieval-augmented generation lets AI answer from your own documents, not memory. See how RAG works in AI voice agents and why it can reduce wrong answers.
Qu’est-ce que le contact de la partie droite (RPC) ?
Le right party contact (RPC) consiste à joindre et vérifier la personne visée par le dossier. Formule du taux de RPC et moyens de l'améliorer.
What Is Role-Based Access Control (RBAC)?
Role-based access control (RBAC) grants permissions to roles, then roles to users. Learn how RBAC works in a call center for agents, supervisors and admins.
What Is RTP (Real-time Transport Protocol)?
RTP carries the actual voice in a VoIP call, while SIP only sets it up. See what RTP does, how RTCP reports loss and jitter, and why one-way audio happens.
Qu’est-ce que le niveau de service (SL) et le SLA ?
Le niveau de service est la part de contacts répondus dans un délai cible. Formule, signification du 80/20, différence avec un SLA, et biais possibles.
Prêt à voir DialerBee en action ?
Demandez une démo en direct de 15 minutes, ou démarrez un essai gratuit de 14 jours — sans diaporama, sans engagement. Les appels démarrent dès que votre propre trunk SIP ou PBX est connecté et approuvé.
Essai gratuit de 14 jours · crédit IA gratuit · sans carte · 11 langues · BYOC · contrôles qui contribuent à la conformité