🔒 Securing AI Systems

When your company deploys an AI system, that system becomes something you have to defend — a new kind of attack surface. This discipline is often called 'AI security' or 'ML security,' and it borrows from traditional security while adding brand-new concerns.

The Layers You Have to Protect

1. 1.The training data — poisoned or leaked data corrupts or exposes the model (guard the supply chain)
2. 2.The model itself — protect against theft, inversion, and adversarial inputs
3. 3.The application around it — prompt injection, insecure output handling, over-broad permissions
4. 4.The infrastructure — the servers, APIs, and pipelines (classic security still fully applies)
5. 5.The humans — the people who use and trust the AI's output (social layer)

Practical Defenses for LLM Applications

• •Never fully trust model output — validate and sanitize it before using it in code, queries, or commands
• •Least privilege for AI agents — an AI connected to tools should only access what it truly needs
• •Keep a human in the loop — for consequential actions (payments, deletions), require human approval
• •Filter inputs and outputs — screen for injection attempts and sensitive data leakage
• •Don't put secrets in prompts — anything in the context window can potentially leak out
• •Log and monitor — treat AI interactions as security-relevant events worth recording