# AI Security Review vs classic penetration test

**Answer first:** a classic penetration test exercises network, web, API, and identity controls. An [AI Security Review](/ai-security) focuses on the **AI cybersecurity** of one LLM app, RAG assistant, or agent: untrusted content in prompts and indexes, retrieval permissions, and actions the model can trigger. It is **not a substitute** for infrastructure or application penetration testing.

| | AI Security Review | Classic penetration test |
| --- | --- | --- |
| Surface | Prompt paths, retrieval, tools, rendered model output | Hosts, apps, APIs, auth, config, often broader |
| Typical findings | Indirect injection via docs/logs, cross-role retrieval, secret leakage in answers, tool abuse, markdown/exfil channels | SQLi, auth bypass, misconfig, exposed services, classic OWASP Web/API issues |
| Best when | An assistant already reaches sensitive data or tools | You need assurance on the stack around (and beyond) the AI feature |
| Arviser role | Scoped AI Security Review after intake | We do **not** sell full infra pentest as a product line |

## When you need both

Many LLM apps sit behind a normal web or API surface. A pentest can miss retrieval ACL failures and document-borne injection; an AI Security Review can miss a misconfigured API gateway. For high-stakes launches, schedule both with clear scope boundaries.

## What “AI cybersecurity” means here

We use **AI cybersecurity** as a plain-language synonym for securing LLM applications that read data and take actions — not as a claim to replace enterprise cybersecurity programs, MDR, or certification audits.

## Related

- [AI Security Review vs LLM Quality Audit](/ai-security/vs-quality-audit)
- [OWASP LLM Top 10 2026 — what we check](/guides/owasp-llm-top-10-2026)
- [Indirect prompt injection](/guides/indirect-prompt-injection)
- [Lethal trifecta](/guides/lethal-trifecta)

**Next:** [AI Security Review](/ai-security#scope) · [LLM security lite (CI slice)](/services/llm-security-review) · [All guides](/guides)
