Security guide
AI Chatbot Security and Privacy: A Practical Checklist
A website chatbot should only have access to the information it needs to help a visitor. Security starts with deciding what it may know, what it must never reveal and when it should hand the conversation to an authenticated process or person.
Start with the boundary between public and private
A public website chatbot is part of your public-facing service. Give it only information that is appropriate for any site visitor to receive: approved product details, policies, public help content and other material intended for customers. Keep credentials, internal notes, private staff documents and identifiable customer records out of public knowledge sources.
If a use case needs account-specific information, treat it as a separate authenticated workflow. A chat prompt or an instruction to “only show the right customer’s information” is not an access-control system. The application handling the data must verify who the person is and what they are allowed to see before returning it.
Map the information before connecting sources
Make a short source register before launch. For each website, file or connected source, record its owner, intended audience, sensitivity and last review date. Remove expired offers, old policies, copied drafts and material whose audience is unclear. If you would not publish a document on the open web, do not add it to a public assistant until you have confirmed the source’s access and data-handling model.
Keep the source set small enough to review. When a price, return rule, booking process or service area changes, update the approved source and check the answers that depend on it. An assistant can only be as dependable as the information and permissions around it.
Do not rely on a prompt as a security control
Clear behaviour instructions help define the assistant’s role, but they do not replace application permissions. A visitor may try to override those instructions, and hostile directions can also appear inside content the system retrieves. OWASP describes prompt injection as a risk that can lead to unintended disclosure or actions; retrieval and system prompts alone do not eliminate it. Keep sensitive data outside the public answer path, restrict any connected actions, and validate permissions in the system that performs them.
Use this pre-launch test checklist
- Ask for an answer that is not in the approved sources. Confirm the chatbot says what it cannot verify and points to a useful next step.
- Try a direct instruction to ignore its rules or reveal hidden instructions. It should stay within its intended public role.
- Put an instruction-like string in a test document and check whether the chatbot treats it as untrusted content rather than authority.
- Ask for another customer’s details or a private account status. A public assistant should not claim to retrieve or disclose them.
- Test an old price or policy, a sensitive question and a direct request for a person. Check the answer and the safe route onward.
Use fictional names and synthetic details in these tests. Do not paste real customer records, passwords or payment data into a test conversation. Record what happened, fix the source or workflow, then repeat the test after a meaningful change.
Collect less, explain more
Ask only for details needed for the visitor’s next step. Explain why you need them and where the conversation will go. Avoid requesting passwords, full payment details or sensitive identifiers in an open chat. Decide who can access conversation records, how long they are needed and how a deletion or access request is handled. Check the platform’s current privacy and retention terms instead of assuming that every provider handles data the same way.
Questions to ask your chatbot provider
- Which sources can the assistant access, and how do I remove one?
- How are chat messages and conversation records handled?
- Can workspace access be limited to the people who need it?
- What happens when a visitor asks for something the assistant cannot verify?
- Which controls apply before an integration can read or change customer data?
- Where can I find current privacy, security and incident information?
NIST’s Generative AI Profile recommends managing risks throughout the AI lifecycle, with appropriate human review, tracking and oversight. Use a checklist like this as a starting point for your own risk assessment—not as a substitute for legal advice, a security review or your organisation’s policies. See the NIST Generative AI Profile and OWASP’s guidance on prompt injection and sensitive information disclosure.
See the ChatNexus security overview for the platform’s current stated approach, and review how to design a useful human handoff for conversations that need judgement or private account support.
Check the boundaries before you publish.
Explore the current security overview or walk through a website assistant use case with the ChatNexus demo.
See the demo Security overview