Prompt injection is a vulnerability in which inputs or injected content change the behaviour or output of a language model in unintended ways.
According to the OWASP Gen AI Security Project, manipulated input can take effect even when nobody can perceive it. The BSI's study on generative AI models distinguishes two routes. In the direct route, users modify their input to bypass safety mechanisms or make the model break out of its predefined role. Depending on where those mechanisms sit, the study calls this jailbreaking or prompt Prompt A prompt is the input a user gives a generative AI system, which then responds with an output. Read in the glossary → injection. In the indirect route, the manipulation hides in unchecked third-party sources such as websites or documents.
According to the study, if a language model Large language model A large language model (LLM) is an AI model, trained on large amounts of text, that processes text inputs called prompts and generates text outputs from them. Read in the glossary → application has execution and access capabilities and can act autonomously on the model's output, unauthorised actions can follow. Examples are unwanted purchases or e-mails sent and deleted. Such an application could be an AI agent AI agent An AI agent is a software system, built on a large language model, that carries out actions on its own and uses interfaces to other programs to reach its goal. Read in the glossary → with access to orders or a mailbox. According to the BSI's 2025 report, there is currently no single highly effective measure against indirect prompt injection, so several are combined in practice. The generative AI AI Artificial intelligence (AI) covers methods that let a computer perform tasks which would require intelligence if a person did them. Read in the glossary → study names one option: making critical actions depend on the user's explicit consent, for example via a confirmation button.
The AI agents we build produce drafts and previews; nothing is published until a person has approved it. In a knowledge platform we built for pharma logistics Pharma logistics Pharma logistics is the term we use for storing, transporting and delivering medicinal products and other pharma-sector goods, such as promotional material and supplies for patients. Read in the glossary → , the chat assistant's own access key reaches only content released to all signed-in users.
Where the term appears
See also
-
AI agent
An AI agent is a software system, built on a large language model, that carries out actions on its own and uses interfaces to other programs to reach its goal.
E-commerce
-
RAG
Retrieval-augmented generation (RAG) is a method that lets a generative AI model draw on documents held in a knowledge base and use them to enrich its output.
E-commerce
-
Prompt
A prompt is the input a user gives a generative AI system, which then responds with an output.
E-commerce
-
Large language model
A large language model (LLM) is an AI model, trained on large amounts of text, that processes text inputs called prompts and generates text outputs from them.
E-commerce
Sources
-
LLM01:2025 Prompt Injection (opens in a new tab)
OWASP Gen AI Security Project (OWASP Foundation) accessed
-
Generative KI-Modelle: Chancen und Risiken für Industrie und Behörden, Version 2.0 vom 17.01.2025 (opens in a new tab)
Bundesamt für Sicherheit in der Informationstechnik (BSI) accessed
-
Die Lage der IT-Sicherheit in Deutschland, Kapitel Resiliente Gestaltung von LLM-Systemen (opens in a new tab)
Bundesamt für Sicherheit in der Informationstechnik (BSI) accessed