03 / AI
LLM & AI application testing
Chatbots, copilots and AI agents tested for prompt injection, data leaks and tools being misused.
Who it's for
Companies that have added an LLM-powered feature, such as a support bot, an assistant or an agent that can take actions, and want to know what it can be tricked into doing before a customer finds out.
Adding an LLM to your product adds a new kind of input: text that can change what your software does. We test what your AI feature can be talked into revealing or doing, and how to stop that without losing the feature.
What gets tested
- Direct and indirect prompt injection, including through documents, web pages and email
- Leaking the system prompt, other users' data or connected documents
- Misuse of the tools and actions the model is allowed to call
- Unsafe handling of model output in your app (XSS, SQL, command injection)
- Access control in retrieval (RAG): who can see which documents
- Cost and abuse controls, such as rate limits and runaway agent loops
What you get
- A report covering the risks your customers and investors ask about, with technical findings for the builders
- Example prompts that reproduce each finding, and a practical fix
- A call to go through the results with the people building the feature
- A free retest once you've fixed things, with an updated report
Typical sizes
| Example | Days |
|---|---|
| Customer support chatbot answering from your help docs | 1–2 |
| Assistant with access to user data and a few tools | 3–4 |
| Agent that can take actions across several systems | 5–8 |
At $1,200 per day. Examples only; every job is quoted after a free scoping call.
Questions
Which models and providers do you test?
Any. OpenAI, Anthropic, Google, open-source or fine-tuned. The testing targets your application around the model, because that's where the risk is.
Isn't this the model provider's job?
Providers secure the model. What it's connected to, what it can see and what it's allowed to do are your team's decisions, and that's what gets tested.
Do you follow a standard?
The test covers the parts of the OWASP Top 10 for LLM Applications that apply to a running product, such as prompt injection, data leakage, excessive agency and output handling. The report is organised by what we found and how to fix it, and if your auditor needs findings tagged against the list, ask and we'll add it.
Get a quote for this test
A few questions about what you've built, and you'll have a fixed quote after a short call.