DIGITAL SECURITY GUIDE
AI Prompt Injection and Data Leakage: The Basics
Digital security for AI workflows includes understanding which information enters a system and which instructions it should trust. AI workflows can process instructions and large amounts of context. Understanding what information enters a workflow—and which instructions should be trusted—is an important part of using these systems safely.
What is prompt injection?
Prompt injection is an attempt to influence an AI system by placing instructions in content the system reads. That content might come from a user prompt, a document, a webpage, or another external source. The risk is that an AI application may treat untrusted content as instructions rather than data.
What does data leakage mean?
Data leakage occurs when information reaches someone or somewhere it should not. In AI workflows, this can happen when a user pastes confidential details into an unapproved service, an application exposes sensitive context, or an integration handles data too broadly.
Reduce avoidable exposure
- Do not paste passwords, API keys, or access tokens into prompts.
- Remove personal and customer information unless it is essential and approved.
- Separate trusted system instructions from untrusted documents and external content in AI applications.
- Limit the data and permissions available to connected tools.
- Review outputs and logs for accidental disclosure.
Use layered safeguards
No single prompt filter can prevent every injection or disclosure. Combine careful data handling, access controls, approved AI services, secure integration design, monitoring, and human review. Pattern scanners can identify some common secrets, but they do not understand every context or inspect every type of content.
For a practical prompt review, try the AI Data Protection scanner and read its limitations before relying on its results.
For a separate check of common sensitive text patterns, try the AI prompt scanner. For organization-wide practices, read AI data protection for small teams.
For everyday prompt hygiene, see how to protect sensitive data in AI prompts, then use the AI prompt scanner as one additional review step.
← All AI security guides