AI agents and chatbots
AI agent and chatbot security audit
Your AI agent reads emails, documents, web pages and chat messages from people you do not know. We check whether one of them can make it move money, leak client data or make a promise you never approved.
From AED 35,000, excl. 5% VAT
Get the sample report Ask for a quote Book a call · WhatsApp
What we try to make your agent do
- Send your data, or a client's data, to an address the attacker controls.
- Use one of its tools for something nobody asked: send an email, change a record, start a payment.
- Make a promise or a commitment in your company's name.
- Reveal its instructions, the secrets in them, or another user's data.
- Tell your staff that it did something it did not do.
We test your agent with the same tools and access it has in daily use.
What we found in our own lab
We built a test email assistant with one tool: send an email. Then we sent it 50 attack emails mixed with 10 normal ones, on five AI models. With no defense, 38 of 250 tries made it call its send-email tool to email the attacker from its owner's mailbox. On one model, 16 of its 50 tries did. The lab records each tool call and never runs it, so no email was sent.
We then tested three fixes, each on its own. They brought the count to 0, 0 and 1 of 250, and every model still summarized the 10 normal emails. These numbers describe our test setup only. In an audit, the same AI security testing runs on your own agent, with its tools and its data.
Get the sample report
The report of our lab test: each finding with its evidence and its fix, and what was not tested. Leave your work email and the PDF link appears here.
Thank you. Download the sample report (PDF)
Prompt injection, AI red teaming and LLM security: what we test
This AI security audit is a penetration test of your AI agent or chatbot, also called an AI pentest, AI red teaming or LLM penetration testing.
- Prompt injection, typed by a user or hidden in what the agent reads: emails, files, web pages, tool results.
- Tool misuse and excessive agency: which tools the agent can call, with which rights, and what an attacker can make it do with them.
- Data leaks: the system prompt and the secrets in it, other users' data, and data sent out through links or images.
- Output handling: what your app does with the agent's answer, such as links, HTML or commands.
Each finding is mapped to the OWASP Top 10 for LLM Applications 2025 and the OWASP Top 10 for Agentic Applications 2026.
What you receive
A written report. Each finding says what an attacker can do, how serious it is, the evidence, and how to fix it. The report also says what was tested and what was not.
See the sample report, built from our own lab tests.
A-1 An email from a stranger makes the assistant call its send-email tool from the owner's mailbox
- What an attacker can do
- Send one email to the owner. When the owner asks for a summary, the assistant calls its mail tool to send the attacker's text to the attacker's address, in the owner's name.
- Evidence
- With no defense, 38 of 250 attack tries made the assistant call its mail tool.
- Fix
- Remove the mail tool when the task is a summary.
- Retest result
- Tool removed: 0 of 250. Marker lines: 0 of 250.
How it works
- A call to agree what is tested, and the price.
- You sign a written authorization. No test starts without it.
- We run the test.
- You receive the report.
Ask for a quote
Tell us what you want tested. We will reply by email.
Thank you, your request is in. We will reply by email.
Published work
Related
- Your app itself was built with an AI tool? See the AI-built app security audit.
- All prices · Who runs the tests · Facts about ARELION