Most pentests miss AI vulnerabilities. I test both.

A traditional pentest covers networks, web apps, and APIs. You need that. But if your product runs an LLM or an agent workflow, a classic pentest misses a whole class of bugs it was never built to find.

Two threat models

Traditional surface: auth bypass, injection, broken access control, misconfigured infrastructure, exposed APIs. Well-mapped, and still where plenty of real risk sits.

AI surface: prompt injection, data leakage through model outputs, insecure tool and function calling, RAG poisoning, guardrail bypass, PII exposure. A standard web-app checklist has nothing for these, because the vulnerable component is the model and the way you wired it in.

Test one surface and you get a false sense of security. The report comes back clean while the newest part of your product, the AI, goes untested.

Full coverage

• The web, app, API, and network testing you expect from any pentest.
• Prompt injection and jailbreak testing against your specific prompts and tools.
• Data-flow analysis: what the model can reach, and what can leak back out.
• Tool and function-calling authorization review.
• RAG and retrieval security if you pull in external content.

One provider for both

Split the work, classic testing to one firm and AI testing to another, and you get gaps at the seams. Your AI features connect to your existing infrastructure at those seams, which is where a lot of the interesting bugs hide. Testing both together catches the bugs that exist only because the two systems talk to each other.

If you're building an AI product, that combined coverage is the point of what I do.