Securing AI &
LLM Applications
Generative AI created a brand-new attack surface that traditional controls don't see. Here's what's different, the risks that matter, and the defenses that actually hold.
Classic appsec assumes code and data are separate. Large language models break that assumption: instructions and data arrive in the same channel — natural-language text. An attacker can smuggle commands inside content the model reads (a web page, a document, an email), and the model may follow them. Add non-deterministic outputs, third-party model APIs, and employees pasting sensitive data into chatbots, and you have risks that firewalls and WAFs were never designed to catch.
The OWASP Top 10 for LLM Applications is the best shared vocabulary. The ones that bite most teams:
Malicious instructions hidden in user input or in content the model retrieves (indirect injection) override your system prompt — the #1 LLM risk.
The model reveals secrets, PII or proprietary data — from its context, training data, or connected tools.
App code trusts model output and passes it into a shell, SQL query, or browser — turning a hallucination into RCE or XSS.
An AI agent has more tools/permissions than it needs, so a successful injection can take real, damaging actions.
Poisoned training data, tampered models from public hubs, or vulnerable AI libraries.
Cost and denial-of-wallet attacks that hammer expensive model calls.
Before you secure the AI apps you built, find the ones your people are using. Employees paste source code, customer data and strategy docs into consumer chatbots and browser extensions every day. This shadow AI is often the biggest real-world leak — and it's invisible unless you're looking for it at the endpoint and network.
An LLM firewall sits in the path of AI traffic and inspects it in real time — much like a WAF for the model layer. A good one will: