AI/LLM Security Assessment
If your chatbot can read your files or send emails, someone will eventually try to talk it into doing the wrong thing. We test it before they do.
Why AI needs its own kind of testing
Ordinary software does what its code says. AI systems do what their instructions say, and anyone who can get text in front of them can try to change those instructions. A cleverly worded message, or a hidden line inside a document, email or web page the AI reads, can make it leak customer data, reveal its private instructions, or take actions nobody approved.
Prompt injection has been the number one risk on the OWASP Top 10 for LLM Applications in every edition so far. The more your AI is connected to, such as your documents, your inbox or your systems, the more damage a successful attack can do.
What we test
- Prompt injection, both direct from users and indirect from documents, emails and web pages the AI reads
- Data leakage: whether the AI can be pushed into revealing customer data, internal documents, keys or its hidden instructions
- Too much power: whether an agent can be tricked into sending emails, changing records, moving money or running code
- Unsafe output: AI responses that end up running as code in your website or other systems
- Poisoned knowledge: whether someone can plant false or malicious content that your AI will trust and repeat
- AI supply chain: third-party models, plugins, MCP servers and libraries
- The application around it: logins, access rules, APIs, rate limits, and abuse that could run up your AI bill
- Guardrail bypass: attempts to get around your content and safety rules
Packages
| AI Chatbot Check | AI Application Assessment | AI Agent Assessment | |
|---|---|---|---|
| Good for | A simple website or support chatbot | An assistant that uses your documents or data | Agents that use tools, APIs or MCP servers |
| Typical scope | One chatbot, no tool access | Document search (RAG) and up to 3 integrations | Multi-step workflows and actions in other systems |
| Typical length | About 2 days | About 4 days | Agreed during scoping |
| Where we test | Usually the live chatbot, under agreed rules | A test copy, then a careful check on live | A test copy, then a careful check on live |
| Price | From $1,990 | From $4,900 | From $8,900 |
| Ask about this | Ask about this | Ask about this |
No test copy? We can test the live system on its own. It takes more care, so it is priced a little higher.
What you get
- A plain-language summary of the business risk
- Each finding with proof, a severity rating, and a mapping to the OWASP LLM Top 10 and MITRE ATLAS
- A prioritised fix plan your developers can work from
- A walkthrough call with your team, and a free retest within 30 days
Test copy or live system?
A test copy (often called staging) is a version of your AI that real customers never touch, so we can push hard without any risk. The trade-off is that a copy is rarely identical to the real thing.
Testing your live system shows exactly what an attacker would see. We only do it under written rules: agreed time windows, limited test accounts, spending caps, no destructive actions, the least possible contact with real data, and a person on your side who can pause us at any time.
Use AI tools but haven't built your own? Your main risk is staff putting company data into public AI tools. See our AI Usage Policy service.
Want to build your own AI tool? We build them too, securely. Tell us your idea.
Common questions
Do you need access to our AI provider account?
Usually not. We test through the same interfaces your users or staff use, plus any test accounts you give us. If a configuration review is in scope, read-only access is enough.
Will testing affect our customers?
Not if we test a copy. If we test the live system, we work in agreed windows with limits on spending and actions, and we stop as soon as we've proved a problem exists.
Which AI platforms can you test?
Any chatbot, assistant or agent built on major model providers or open-source models, including custom builds, low-code AI tools and MCP-based agents.