An 8-point AI security checklist
| Risk | What to check |
|---|---|
| 1. Data leakage | Are prompts and documents sent to an external model provider? Will they be used to train models? |
| 2. Loss of access control | Is the knowledge base the AI can access configured by role? Does revoking access take effect immediately? |
| 3. Prompt injection | Is imported document and web content sanitised? Are user prompts checked? |
| 4. Misuse of external tools | Which external tools can the AI call? Do they require approval, and can they be restricted to read-only? |
| 5. Personal data | Can personal data be masked on import? Can you set conversation retention periods and delete records? |
| 6. Wrong answers with no source | Do answers cite their source documents? When the evidence is insufficient, does the AI say so clearly? |
| 7. No activity logs | Is there a record of who used it, which model was used and how the request was processed? |
| 8. Runaway usage | Can you set quotas, receive notifications near the limit and trace where costs come from? |
What is prompt injection?
Prompt injection hides malicious instructions in user input, documents or web content so that the AI ignores its original rules, for example by leaking data or performing operations it should not. The risk is higher when the AI can read external content and call tools. Common countermeasures include sanitising imported content, checking prompts, restricting tool permissions and keeping activity logs.
Frameworks you can refer to
- OWASP Top 10 for LLM Applications: a summary of common security risks in large language model applications.
- The Executive Yuan's guidelines for the use of generative AI by government agencies: principles for generative AI use in Taiwan's public sector, which private enterprises can also refer to.
- ISO/IEC 27001: AI systems can be brought into the scope of an existing information security management system.
How NOVARK helps
In NOVARK, AI governance is enforced centrally by the platform. Access control is set by role and by individual, and revoking access takes effect immediately. External tools must be approved first, and each time before use the platform confirms their content has not been changed. Knowledge answers cite the document, page and paragraph. A single trace ID links user actions, model calls and execution logs. Fully on-premises deployment and personal data masking are supported, and monthly usage quotas can be set per system and per user, with usage paused automatically when the limit is reached.
NOVARK Audit
NOVARK DocFAQ
Is using a cloud AI model always unsafe?
Not necessarily. The key is to confirm which data will be sent externally and how the provider handles it, and to decide which work can use cloud models based on how sensitive the data is.
What are the risks of employees using free AI tools on their own?
Company data may be pasted into external services, with no access control and no usage records. Provide managed enterprise AI tools and set clear usage policies.
This article is general guidance. Adapt the approach to your organisation's size, industry regulations and data. NOVARK is an enterprise AI agent platform.