AI security includes protecting data and controlling what a model-connected application can do. External documents and web pages may contain instructions that attempt to redirect the system, even when the application was asked only to summarize their content.
Treat retrieved material as untrusted input and enforce permissions outside the model. Limit access to the records and tools required for the task. Validate proposed actions and outputs with application controls rather than assume a system prompt will prevent every misuse.
Test direct and indirect prompt injection, inappropriate disclosure and unauthorized tool use. OWASP explains that retrieval and fine-tuning do not fully eliminate prompt injection. The effect of an attack depends partly on the privileges and downstream actions available to the application.
Log relevant security events without unnecessarily exposing sensitive content. Provide a way to revoke access and stop the workflow. Review the controls after changes to tools, data sources or models, and require appropriate authorization for consequential actions.
OWASP prompt injection risk and mitigations
Open Sources Used
This page uses open and institutional references as a frame; the final decision still belongs to the company record, threshold and owner.
Related Articles
Reading adjacent decision areas keeps the topic from becoming an isolated note.