Ryvx

If your website has a chatbot or an AI agent, it can be attacked in ways your normal website test does not cover. The most common is prompt injection: someone types, or hides in a page the bot reads, instructions that make it ignore its own rules.

What can go wrong

A chatbot that can look up orders might be talked into showing another customer's order. An agent that can send emails or issue refunds might be pushed into doing it for the wrong person. A bot might reveal its hidden instructions, internal links or data it was never meant to share. These are the risks listed in the OWASP Top 10 for large language model applications.

What a good test covers

A useful test starts with your limits: what the bot must never do. It then sends real attacks, both typed directly and hidden in content the bot reads, and reports what was tried, what got past your limits, and what was not tested. Be wary of any report that only lists passes with no record of the attempts behind them.

How Ryvx tests chatbots

Chatbot and agent testing is part of a Ryvx Launch Security Review, £495 per site, alongside the test of the website itself. You tell us what the bot must never do. We run 11 attack categories against the live chatbot by default, covering 8 of the 10 OWASP risks for AI applications, including prompt injection, sensitive data disclosure, excessive agency and system prompt leakage.

The other two, supply chain and training data poisoning, cannot be tested by talking to a chatbot. The report says so, with the reason, rather than leaving a blank. As with the rest of the review, a finding only counts when it comes with proof, and a person checks every report before you see it.