If you have put a chatbot, an AI agent or an AI-powered feature in front of customers or staff, it can be attacked in ways a normal website can't. Kavis Network tests these AI systems the way an attacker would, with your written permission, and gives you a prioritised report of what to fix.
AI security testing vs AI-assisted pentesting
These sound alike but are different things:
| AI security testing (this service) | AI-assisted pentesting | |
|---|---|---|
| What is tested | The AI system itself: agents, LLM apps, RAG chatbots | Ordinary web and SaaS applications |
| Role of AI | AI is the target | AI is the tester's tool |
| Typical issues | Prompt injection, data leakage through answers, agents misusing tools | Broken access control, IDOR, injection, misconfiguration |
| Where to learn it | AI Security Course | SaaS Product Pentesting with AI |
What we test
- Prompt injection: direct attempts by users, and indirect instructions hidden in documents, web pages or emails the AI reads.
- Data leakage in RAG chatbots: whether one user can retrieve another customer's data or internal documents, and whether access rules apply before retrieval.
- System prompt exposure: whether the bot can be made to reveal its instructions, keys or internal details.
- Agents and tools: excessive permissions, risky actions without approval, and misuse of connected tools.
- Output handling: unsafe links, scripts or images in answers that could leak data or harm users.
- Abuse and cost: missing rate limits, and usage that could run up your AI bill.
How an engagement works
- Scoping call. We agree what is in scope, the test environment and the dates.
- Written authorisation. We test only systems you own or have permission to test, as set out in writing.
- Testing. Hands-on attack attempts against the agreed targets.
- Report. Findings ranked by risk, with evidence and clear fixes your developers can act on.
- Walkthrough. We go through the report with your team. A retest after fixes can be agreed in the proposal.
What to have ready
- The chatbot, agent or app to test, ideally a test environment with realistic but non-sensitive data.
- Test accounts for different user roles, so we can check one user can't see another's data.
- A list of the tools, APIs and data sources the AI can reach.
- A contact who can approve testing and answer questions during the test.
What is and isn't included
| Included | Not included unless agreed in the proposal |
|---|---|
| Testing of the agreed AI components | A full pentest of your wider network or infrastructure |
| Prioritised report with fixes | Implementing the fixes in your code |
| Report walkthrough | Retesting after fixes |
Who this is suitable for
- Businesses about to launch, or already running, a customer-facing chatbot, AI agent or RAG search.
- Teams whose AI can read private data or take actions such as sending emails or updating records.
When another approach may be better
If your AI tool only drafts text that a person always reviews, and it can't reach private data or take actions, a short review using our checklist-style articles may be enough. Start with 7 checks before you launch a RAG chatbot.
Pricing
Pricing depends on scope and is set out in a written proposal.