AI security 101: Protect data and actions

How do you protect systems using AI?

An AI assistant reading documents has one level of risk.

An AI agent changing customer records or issuing payments has another.

Your security controls should reflect the data AI accesses and the actions AI performs.

Three risks deserve attention:

1. Prompt injection: instructions hidden in a document, website, or message attempt to redirect the AI. For example, a malicious email tells an agent to send customer information to an outside address.

2. Data leakage: sensitive information reaches an unauthorized person or service through prompts, retrieved documents, or generated responses.

3. Excessive permissions: an AI system receives more access than its task requires. An agent checking an invoice should not automatically have permission to approve payment.

How do you reduce these risks?

• Limit access. Give each AI system only the data, tools, and permissions required for its work.
• Protect sensitive data. Define which information approved AI services are allowed to process.
• Treat external content as untrusted. Enforce access and action controls outside the model. A prompt asking AI to behave securely is insufficient.
• Require approval for consequential actions. Payments, record deletion, and permission changes need stronger controls.
• Monitor execution. Record tool use, investigate exceptions, and give your team the ability to stop the system.

Vendors offer a wide range of solutions addressing different aspects of AI security.

Examples include Cisco AI Defense, Palo Alto Networks Prisma AIRS, Microsoft Purview, Check Point Software AI Network Firewall, Fortinet FortiAIGate and Cloudflare AI Security for Apps.

Each solution has different capabilities.

Match your selection to your applications, data, and specific risks.

My takeawaygive AI a defined job, limited authority, and accountable oversight.

Do your AI systems have permissions matched to their tasks?

Learn more about the solutions:

• Cisco AI Defense: lnkd.in/ec_uz6_b
• Palo Alto Networks Prisma AIRS: lnkd.in/eDFbzceR
• Microsoft Purview for AI: lnkd.in/exsCWPPE
• Check Point AI Network Firewall: lnkd.in/eu6rm2E3
• Cloudflare AI Security for Apps: lnkd.in/eSKdXy66
• Fortinet FortiAIGate: lnkd.in/eEs7Uga4