AI Security / 2026 Essentials
Keep Sensitive Data Out of Unapproved AI Workflows
Understand what an AI service receives and what its answers might expose.
Environment and inputs: One approved AI workflow and synthetic input containing a fake identifier and fake secret for checking outbound fields.
AI applications can disclose personal information, credentials, or confidential business material through their inputs and outputs. A system prompt saying “keep this private” is not an access-control boundary. Minimise the information supplied and review the complete application path, including connected data sources and generated responses.
What matters
- Classify information before sending it to an external AI service.
- Remove unnecessary identifiers and secrets from prompts, examples, and reference material.
- Verify vendor settings and enforce access controls in the application, not only in instructions.
Practical example: A troubleshooting question includes a sanitised error message instead of a full configuration containing production credentials.
One action today: Inspect one real AI workflow and remove sensitive fields that are not needed for the task.