SteadySec Your retained security partner

Retainer capacity is full until mid-2027.Fixed-scope projects available on request.

AI Application & Agent Pentesting

Your AI answers questions, and increasingly it takes actions.
We test what an attacker can make it do.

  • OWASP AI Testing Guide (AITG-APP)
  • EU AI Act mapping
  • fixed price after scoping
  • from €2,900

This assessment can be booked on its own, or scheduled for you inside the Retained Security Partner retainer at the point where it adds the most.

The Challenge

Why an AI application needs its own test

Your customers ask

Buyers whose data your AI touches ask what testing it has had. One report answers every questionnaire.

security questionnaires · vendor due diligence

Your ISMS already says so

ISO 27001 asks you to test the systems in your scope, and so does NIS2 where it applies. An AI application handling customer data is one of them.

ISO 27001 A.8.8 / A.8.29 · NIS2 Art. 21(2)(e)

The attack is a sentence

An attacker steers the AI with plain text in a chat, a document or an email it reads. Testing for it takes AI-specific cases, which a web pentest scope rarely includes.

EU AI Act · mapped where it applies

One family of attack: injection

Four ways untrusted text reaches the model, the one place it all arrives, and what it costs.

Text the application does not control

User prompt

Direct prompt injection

Retrieved documents

Indirect injection in documents

Tool and API responses

Injection through hijacked tool output

Sub-agent messages

Injection carried between agents

All of it arrives here

One context window

Every source read with the same trust

The attacker's instruction runs

  • Another customer's records returned The retrieval boundary fails
  • A refund issued, a record changed A tool runs for the wrong person
  • An internal API reached from outside Model output executed downstream

A sentence inside a document your customer uploaded is read with the same weight as an instruction from you. That is why each entry point has to be tested where it joins the flow.

Scope

What's in scope

What your AI can do decides which of these apply. We confirm that during scoping.

Data exposure

Cross-customer and cross-session retrieval, and what it gives away about its own prompts, policies and tools.

Poisoned content

Planted documents or agent memory surviving into a later session with a different user.

Privilege escalation

Whether an ordinary user can reach an action meant for an administrator. If your AI can act.

Transaction abuse

Refunds, record changes, outbound messages: who can trigger a tool call, and with what arguments. If your AI can act.

Internal access

Model output reaching an internal API, a database query or a shell in a form that gets executed. If your AI can act.

Cross-agent escalation

A delegated action exceeding the permissions of the agent that asked for it. If agents delegate.

Application layer. The guide's infrastructure, data-pipeline and model categories sit outside this assessment. Model-layer work on a model you trained yourself belongs with a specialist ML security team.

Findings mapped to: OWASP AI Testing Guide - AITG-APP EU AI Act / NIS2 / DORA

Pricing

Indicative pricing

Assistant

Text in, text out, no tools. We test what it discloses, what it renders and whose documents it can reach.

Agent

It calls tools, functions or an MCP server. We test the tool layer as well, where the most serious findings usually are.

Starting from
€2,900
exact quote after scoping
Always included
Coverage of the applicable AITG-APP categories
Prompt injection, jailbreak and guardrail testing
Tool-layer authorisation checks
Findings led by business impact, with reproduction steps
Executive summary for management
EU AI Act / NIS2 regulatory mapping
One verification round after remediation - included
Debrief call with your dev team

What affects the final price

Tools and integrations

Every tool the AI can call is a separate attack path. More than twelve tools moves the test above the standard price.

Tenancy and data

A knowledge base holding per-customer data needs an account in each of two tenants, and doubles the retrieval testing.

Lead Time saves money

Booking at least three weeks ahead lowers the quote. Urgent engagements carry a premium.

Get started

Scope a test

Answer ten questions about what your AI can do and who can reach it, and we come to the first call with a price.

Have questions? See the FAQ →

For regular testing on a monthly fee, see the Retained Security Partner retainer.