# AI Security Review for LLM Apps and Agents

AI cybersecurity of systems that read your documents and take actions. Find what your assistant can expose or do beyond its permissions.

**Scoped after intake · One application · RAG track available · Warsaw**

HTML: [/ai-security](/ai-security) · Markdown: [/ai-security.md](/ai-security.md)

## What we test

- **Prompt injection** — can instructions in a user message, retrieved document, tool response, or log change behaviour?
- **Data access** — can a user retrieve information their role should not see (including cross-tenant chunks)?
- **Agent permissions** — can the agent act outside its role or skip approval?

## RAG track

Sources → ingestion → index → retrieval → prompt → model → tools → rendered answer. Covers indirect injection, retrieval ACL, secrets in the index, output channels, and write-path poisoning.

## Compare

- [vs LLM Quality Audit](/ai-security/vs-quality-audit)
- [vs classic pentest](/ai-security/vs-pentest)
- [Risk catalog R-01–R-27](/guides/ai-security-risks)
- [GitHub catalog (MIT)](https://github.com/nareyko/ai-security-risks)
- [OWASP LLM Top 10 2026 mapping](/guides/owasp-llm-top-10-2026)

## CTA

Request a security review: [/ai-security#scope](/ai-security#scope)

Point-in-time review. Not OWASP certification. Not a substitute for full infra/web pentest.
