Skip to content

// Pentest by target type

Pentest for AI applications

7 minUpdated November 16, 2026By Ler em português

Why a product with AI needs its own test

In an ordinary system, the code decides what happens. In a product with a language model, part of the decision moves to the model, and the model reads text from anywhere: what the user types, an uploaded document, a web page, an e-mail.

That opens flaws a traditional website and API test does not look for. OWASP, the industry reference, has kept a list for this kind of application since 2023. The 2026 edition, published in August, has ten items and is the basis of what RET tests.

RET runs this test: a 25-hour Sprint for R$9,500, within 7 business days. See price and timeline for my case

What the test looks at

In plain words, with the name the 2026 OWASP list gives each risk:

  • someone writes an instruction, in the chat or hidden in a document, and the model starts obeying it instead of your rules (Prompt Injection);
  • the model returns another customer’s data, a company secret or part of what it was trained on (Sensitive Information Disclosure);
  • text the model generates is used unchecked and becomes a command, a database query or code on the page (Improper Output Handling);
  • the agent has more tools and permissions than it needs, and a malicious instruction makes it delete, send or pay (Excessive Agency);
  • internal instructions and context the user should not see leak and reveal a business rule, a tool or a key kept there (Hidden Context Exposure);
  • document search returns another customer’s content or accepts a poisoned document (Vector and Embedding Weaknesses);
  • there is no usage limit, and an attacker runs up model cost or takes the service down (Unbounded Consumption).

What stays the same

Most serious flaws in AI products are still the usual ones: one customer seeing another’s data, weak login, billing that accepts a tampered amount. The test covers both layers, the product and the model, because a hidden instruction only becomes a loss when it meets a permission that should not exist.

Delivery and terms

Executive and technical reports with reproducible evidence and recommended fix, NDA the same day, Sprint within 7 business days, and a second check of your fixes with no limit for 90 days.

Frequently asked questions

Does an API pentest already cover the AI part?

No. An API test checks who may call what. It does not try to make the model disobey its rules, leak internal instructions or misuse a tool.

My product only calls a model provider’s API. Do I need a test?

The provider’s model is not the target. Your product is: what you put in the context, which tools you give the model and what you do with the answer.

How much does it cost?

The 25-hour Sprint costs R$9,500 for one web application, API or mobile app (Android or iOS).

What if the test finds no flaw?

We refund 50% of the price. The other 50% pays for the hours executed, the report of what was tested and the attestation letter.

Keep reading