Security foundations

AI systems should follow the same security principles expected of other enterprise systems.

Data protection

Identify:

• Sensitive information

• Personal information

• Confidential business data

• Restricted datasets

Identity

Every system interaction should have a defined identity.

Access

Use least-privilege permissions.

Monitoring

Record meaningful:

• Access events

• System actions

• Failures

• Permission changes

Prompt and input security

Systems should account for:

• Untrusted inputs

• Malicious instructions

• Unexpected data

• Prompt manipulation

Output validation

AI-generated outputs should not automatically become trusted system actions.

Where appropriate:

AI output
   ↓
Validation
   ↓
Policy check
   ↓
Action

AI output
   ↓
Validation
   ↓
Policy check
   ↓
Action

AI output
   ↓
Validation
   ↓
Policy check
   ↓
Action

AI output
   ↓
Validation
   ↓
Policy check
   ↓
Action

Incident response

Define what happens when:

• Unauthorized access occurs

• Sensitive data is exposed

• An agent behaves unexpectedly

• An integration fails

• A security boundary is breached

Security checklist

• Data classification completed

• Access boundaries defined

• Credentials secured

• Logging enabled

• Output validation implemented

• Incident owner identified

• Security review completed