Is your AI chatbot leaking more than it should?

What we test

• Prompt injection — direct and hidden in documents, web pages or emails the AI reads
• Jailbreaks that bypass your assistant’s rules or tone
• Leaks of hidden system instructions, API keys or internal notes
• Access to other customers’ data through the chat (e.g. orders, tickets, account details)
• Unsafe actions by AI agents that can send emails, update records or call tools
• Retrieval (RAG) issues: documents the AI can see that the user shouldn’t
• Output handling: links, scripts or markup the AI returns that your site then trusts

Why AI features need their own testing

Large language models follow instructions from whatever text they read. That makes them powerful — and easy to manipulate. A chatbot connected to your customer data or business tools can become a way round your normal access controls if it isn’t designed and tested carefully. Our testing is guided by the OWASP Top 10 for Large Language Model Applications, alongside standard web application testing of the page and API the assistant runs on.

How it works

• Apply online and tell us about your AI features
• Prove you own the website with a DNS record or verification file
• We agree the scope, test accounts and testing window with you in writing
• We test, then share a confidential report through your secure portal

Common questions