Free Template · Security Guidelines
AI Prompt Engineering
Guidelines
Security-focused prompt engineering guidelines for enterprise teams. Covers safe prompting practices, data leakage prevention, prompt injection awareness, output validation, and approved prompt templates for common business tasks.
5 sections
security guideline coverage
4 patterns
approved prompt templates
OWASP LLM
top 10 injection risks covered
Free
to use and customise
Why Enterprise Prompt Engineering Guidelines Matter
Employees use AI tools every day to process business information. Without guidance on safe prompting, the default behaviour is to paste whatever data is relevant into the prompt, including PII, confidential client data, and credentials.
#1
Prompt injection is the top LLM security risk
OWASP's LLM Top 10 lists prompt injection as the primary security risk for enterprise AI systems. Without training, employees cannot recognise it.
Data
PII regularly enters AI tool prompts without policy
Without data classification rules for prompts, employees default to including the full context of their work, including personally identifiable data.
GDPR
Sending personal data to AI tools may violate GDPR
Depending on the AI provider's data processing terms, including personal data in prompts may constitute an unlawful international data transfer.
Agents
AI agents amplify injection risk
AI agents that can take actions (send emails, query databases, browse the web) are particularly vulnerable to prompt injection, one injected instruction can hijack the entire task.
The Guidelines
Click each section to expand the guideline content. Share this with your teams and embed it in your AI acceptable use policy.
Safe prompting means providing AI tools with the context they need to do the task, and nothing more. The context window is not a secure container; its contents may be logged, retained, and exposed in future outputs.
Must NEVER be included in any AI prompt:
- Personal Identifiable Information (PII), names, addresses, National Insurance/SSN, dates of birth, financial account numbers
- Credentials, API keys, passwords, authentication tokens, secrets of any kind
- Confidential business information, unpublished financials, M&A activity, pricing strategy, competitive intelligence
- Client or customer data subject to confidentiality agreements or data processing agreements
- Source code for production systems unless the tool is explicitly approved for code use and contains no secrets
- Data classified as Restricted under your data classification policy
Context Window & Memory Considerations
- Assume everything in the context window may be logged by the AI provider, even in enterprise tiers
- Multi-turn conversations accumulate context: earlier messages containing sensitive data remain in context throughout the session
- AI tools with memory or persistent context carry information between sessions, never rely on data being forgotten
- When using RAG (retrieval-augmented generation), data retrieved from external sources is inserted into the context window and subject to the same rules as directly entered data
How to Implement These Guidelines
Follow these five steps to go from guidelines template to organisation-wide enforcement within 30 days.
1
Assess your current AI tool landscape
Identify all AI tools in use, approved and unapproved, and what data employees are entering into prompts today. This baseline reveals the highest-risk behaviours to address first.
2
Define data classification rules for prompts
Map your existing data classification tiers to explicit rules about what can and cannot be included in prompts. Produce clear examples of acceptable and unacceptable prompt patterns.
3
Develop approved prompt templates for common tasks
Work with department heads to create security-reviewed templates for the tasks employees perform most frequently. Pre-built templates reduce the cognitive load of safe prompting.
4
Train employees with examples not rules
Deliver a 30-minute training session focused on concrete examples of safe and unsafe prompts. Include real examples of what prompt injection looks like and how to recognise it.
5
Enforce and monitor with technical controls
Implement technical controls that detect sensitive data in AI prompts in real time. Written guidelines without technical enforcement are aspirational. Monitor for injection patterns automatically.
FAQ
Frequently Asked Questions
What is prompt injection and why does it matter for enterprises?
Prompt injection is an attack where malicious instructions embedded in content cause an AI model to override its intended behaviour when that content is included in a prompt. In an enterprise context, this matters because employees may unknowingly include injected content in prompts (e.g. when asking an AI to summarise an email containing injection instructions); attackers can craft documents that manipulate AI agents into leaking data; and AI systems that act autonomously are particularly vulnerable because injected instructions can redirect their entire task flow.
What data should never be included in an AI prompt?
The following must never be included in any AI prompt, even with approved tools: personal identifiable information (PII) including names, addresses, and financial account numbers; credentials of any kind including API keys, passwords, and authentication tokens; confidential business information including unpublished financials and M&A activity; client data subject to confidentiality agreements; and source code for production systems unless the tool is explicitly approved and the code contains no secrets. When in doubt, anonymise before prompting.
When does AI output require human review before use?
AI output always requires human review before use for: any legal or compliance document, AI output must never be relied upon for legal advice or regulatory submissions; any factual claim for external publication, AI models hallucinate and citations must be verified; any code for production deployment, AI-generated code requires security review; any customer-facing content in regulated contexts; and any decision affecting an individual's employment, credit, or benefits.
How do I test a prompt for security vulnerabilities?
To test a prompt for security vulnerabilities: attempt to override the system prompt by adding instructions at the end of user input; test whether the model reveals its system prompt when asked; for RAG systems, include injection instructions in a test document and check whether they execute; test for data exfiltration by adding instructions to include specific data in the response; and test output filtering by attempting to elicit content that should be blocked. Document all vulnerabilities found and the prompt hardening measures applied.
Get started
Enforce Prompt Guidelines in Real Time
Aona enforces prompt engineering policies in real time, detecting sensitive data in AI prompts before it reaches the model, blocking prompt injection patterns, and logging all AI interactions for compliance evidence.