Services / AI Security Assessment
AI Security Assessment for Chatbots, Assistants, and Agents
An AI security assessment shows how your AI chatbot, assistant, or agent holds up against someone trying to misuse it, or how safely your team uses AI tools with company data. Sanctus tests the way an attacker would and gives you a prioritized plan to fix what we find.
Sound familiar?
- You are adding an AI assistant to your product and want to know it is safe before customers use it.
- A customer or investor asked how you secure your AI features and you don't have a good answer.
- Your staff use ChatGPT, Claude, or Copilot and you aren't sure what company data is going into them.
An assessment gives you evidence you can act on and share, instead of a guess.
AI adds risks a normal security test doesn't cover
An AI model follows instructions written in plain English, and it can't reliably tell your instructions apart from text an attacker slips in. That text can arrive in a chat message, an uploaded document, a web page, or an email the assistant reads for a user.
The model is also only one part of the system. The application around it decides who can see which conversations, which documents the AI can draw on, and which actions it can take. A weakness in any of those can turn a clever message into a data breach.
We help in two ways. An AI product assessment tests the AI features you build or ship. An AI usage review looks at the AI tools your own team uses.
What we test in an AI product
Hidden instructions
Whether a user, or content the AI reads, can slip in instructions that override yours. This is known as prompt injection, and jailbreaking is the version aimed at your safety rules. We test direct attempts in the chat and indirect ones hidden in documents and web pages.
Data leakage
Whether the AI can be coaxed into revealing its hidden setup instructions, internal details, credentials, or information that belongs to your business or other customers.
Separation between users
Whether one customer can reach another's conversation history or shared chats, and whether logins, sessions, and permissions on the chat and history features hold up.
Documents and answers
If the AI answers from your documents, whether it respects who is allowed to see what and whether planted content can change its answers. We also check what happens to its answers afterward, since unchecked output can be used to attack your users or your own software.
Actions and misuse
If the AI can take actions, such as sending messages or changing records, whether it can be talked into doing things you didn't intend, and whether heavy use could run up your costs or take the feature down.
What an AI usage review covers
- Which AI tools are in use, including ones staff signed up for on their own
- What business, customer, and employee data is being entered into them
- What each tool is connected to, such as email, file storage, and code, and whether those connections expose more than intended
- Vendor terms on data retention and model training, and whether your rules match what people really do
You finish with a clear picture of your exposure and a short list of changes that reduce it without banning tools your team finds useful.
Recognized frameworks, careful handling
Product testing follows the OWASP Top 10 for Large Language Model Applications for the AI-specific risks and the OWASP Web Security Testing Guide for the application around it. Automated tools help with coverage, and we verify every finding by hand before it reaches your report.
Testing is controlled and non-destructive. We use dedicated test accounts, an agreed testing window, and agreed rate limits. We never go after other users' data, and if we run into real user data we stop and tell you.
How it works
-
1. Scoping
We agree on which AI features, models, integrations, and data sources are in scope and what you are most worried about. You set up test accounts, we settle the testing window and rate limits, and you get a fixed quote before work starts.
A few days, mostly on your schedule
-
2. Testing
We test by hand, using automated tooling only within the limits we agreed. Anything critical is reported right away instead of waiting for the final report, so your team can start fixing it.
Typically two to four weeks, scheduled around your team
-
3. Reporting and walkthrough
You receive a plain-English report with evidence for every finding and a prioritized plan, then we walk your team through what we found and what to fix first.
Report within a week, then a one to three hour walkthrough
What you receive
- Executive summary you can share with leadership, your board, or a customer asking how you secure your AI features
- Findings report listing each issue with its severity, the evidence behind it, and how to fix it
- Prioritized remediation roadmap so your team fixes the riskiest problems first
- Results walkthrough with the people who built and run the AI feature
Retesting of critical and high severity findings is included, so you can show that the fixes worked.
Pricing
AI security assessments are quoted at a fixed price once we understand what is in scope. The price depends on how many AI features, integrations, and user roles you have, and whether you want a product assessment, a usage review, or both. Contact us for pricing.
If your AI feature runs on cloud infrastructure you also want checked, a cloud security review pairs well with it.
Who it's for
- Software and service companies adding an AI assistant, chatbot, or agent to their product
- Teams about to launch an AI feature who want it tested first
- Organizations whose staff already use AI tools with company data
Assessments are performed by GPEN, GCIH, and GSEC certified consultants. We are independent. We don't build AI products or resell AI tools, so we have no reason to downplay what we find.
AI security assessment questions
What is an AI security assessment?
An AI security assessment is a structured test of an AI feature, such as a chatbot, assistant, or agent, to find ways it can be tricked, made to leak data, or used to harm your users. It covers AI-specific risks and the ordinary application security around the model.
What is prompt injection?
Prompt injection is an attack where someone hides instructions in a message, document, web page, or email so the AI follows them instead of yours. It is the top risk on the OWASP Top 10 for Large Language Model Applications.
How much does an AI security assessment cost?
We scope every assessment up front and quote a fixed price before any work starts. The price depends on how many AI features, integrations, and user roles are in scope. Contact us for pricing and we'll give you a number once we understand what you've built.
How is this different from a penetration test?
A penetration test looks for weaknesses in networks and applications. An AI security assessment covers the same fundamentals for your AI feature, such as logins and separation between users, and adds risks that only exist with language models, such as prompt injection and AI misuse of its tools.
Do you test the AI model itself?
We test how your product behaves, which is what customers and attackers actually interact with. That includes the model's responses and everything wrapped around it. We don't attack your model provider's infrastructure or modify models.
Can you guarantee our AI feature can't be tricked?
No one honestly can. Language models can respond differently to the same input, so a clean result is a strong point-in-time finding and not a guarantee. We tell you what we tried, what worked, and how to reduce the risk, and we retest the serious findings after you fix them.
We don't build AI, we just use it. Is this for us?
Yes. An AI usage review looks at which AI tools your team uses, what data goes into them, and what they are connected to. It's a good fit if your staff use AI assistants and you aren't sure what that means for your data.
How long does an assessment take?
Scoping usually takes a few days, mostly on your schedule. Testing is typically spread over two to four weeks so it fits around your team and any rate limits. The report is ready within a week of testing, and the results walkthrough usually takes one to three hours.
Ready to find out where you stand?
Tell us what's prompting the question. We'll come back with a few questions of our own, not a sales pitch.
Find Out Where You Stand