Skip to content
AI-SNS.IO — CHAT SUPPORTAI-SNS.IO — CHAT SUPPORT
Security & Compliance

Security Checklist for AI Chat Vendors: 12 Questions to Ask

AI SNS CHAT EditorialPublished 7 min read

A practical security checklist for evaluating AI chat vendors: data handling, access control, channels, retention and proof. Use it before you sign.

Choosing an AI chat vendor means handing a third party your customers' questions, your knowledge base and, often, a connection to WhatsApp, Messenger or your website. Marketing pages make every vendor sound safe, so the real work is asking specific questions and checking the answers. This checklist gives you twelve of them, grouped by risk area, with notes on what a good answer looks like and what should make you pause.

Start by mapping what the chatbot will touch

Before you score any vendor, write down what the chat system will see and do. Security requirements follow from that, not from a generic template.

Answer these four questions internally first:

  • What data will customers type in? Names, emails and order numbers are common. Health or payment details should be discouraged outright.
  • What will the bot read? Help-centre articles are low risk. Internal wikis, contracts and shared drives are not.
  • What can the bot do? Answering questions is one thing. Looking up orders, creating tickets or writing to a CRM is another.
  • Where will it run? A website widget, WhatsApp and Messenger each have different identity and webhook risks.

A bot that only answers from public FAQs needs a lighter review than one that can look up customer records. Be honest about which one you are buying.

Identity and access: who can get into the account

Most real incidents start with a stolen login, not a clever attack on the model. Check these first.

1. Two-factor authentication and lockout

Ask whether users can enable authenticator-app (TOTP) two-factor sign-in, and whether administrators can require it. Ask what happens after repeated failed logins. A sensible vendor locks the account for a period instead of allowing unlimited guesses.

2. Roles and permissions

You want more than "admin" and "everyone else". Look for distinct roles (owner, member, support, read-only viewer) and the ability to override permissions per person. A support agent who reads conversations should not be able to delete the knowledge base.

3. Single sign-on, honestly stated

If you need SSO or SAML, ask for a live demonstration, not a roadmap slide. Some vendors list it on a plan comparison without it being available. If the answer is "not yet", you can still proceed, but you should compensate with mandatory 2FA and tight role assignments.

4. An activity log you can read

Ask for a record of who created, edited, deleted or uploaded what, with name, email and time. When something changes in your knowledge base and the bot starts giving wrong answers, this log is how you find out why.

Data handling: where your content goes

5. Encryption and credential storage

Ask how passwords and third-party secrets (API keys, tokens) are stored. Passwords should be hashed with a slow algorithm such as bcrypt, and secrets should be encrypted at rest with a modern cipher such as AES-256-GCM. Ask also whether secrets are redacted from logs, since logs are a common leak path.

6. Retention and deletion

Find out how long conversations are kept, whether you can set a shorter retention window, and what the default is. "Keep forever unless you change it" is a legitimate default, but you need to know it. Also ask how deletion works for an agent, a data source and a whole account, and whether there is a recovery period.

7. Personal data in answers

A chatbot can accidentally repeat personal data from its training material to a stranger. Ask whether the vendor scrubs personal information from knowledge-base passages before showing them to anonymous users.

8. Which AI model providers see your data

Every AI chat vendor sends text to one or more model providers. Ask:

  • Which providers are used (for example Anthropic, OpenAI, Google)?
  • In what order does failover happen if one is down?
  • Can you pin replies to a single provider if your policy requires it?

Failover is good for uptime, but it also means your text can reach more than one company. Your data-processing review should cover each provider in the chain.

  1. 01Customer on web, WhatsApp or Messenger
  2. 02Vendor platform and webhook check
  3. 03Knowledge base lookup
  4. 04AI model provider
Where a customer message travelsIllustrative example

Channels and integrations: the edges of the system

9. Webhook verification

When WhatsApp, Messenger or LINE sends a message to the vendor, the vendor should verify the signature so that nobody can forge incoming traffic. Ask which channels have signature checks, and how outbound webhooks to your own systems are signed. An HMAC signature plus a delivery log lets you verify and debug.

10. API keys and rate limits

If you will use the API, ask how keys are issued. Good practice is that keys are shown once, can be revoked, and are limited to certain plans. Ask for published rate limits so you know what protects the public chat endpoint from abuse.

11. Prompt-injection and abuse controls

Public chat widgets get hostile input. Ask whether filters for prompt injection and abuse are configurable per agent, and whether there are per-device rate limits. A bot that can be talked into ignoring its instructions is a risk, especially if it can call tools.

Proof and paperwork: what the vendor can show you

12. Certifications, stated precisely

Certifications are easy to imply and hard to hold. Ask for the exact wording of what the vendor is certified for, the certificate number where one exists, and the scope. Pay attention to the careful phrases:

  • "Designed with reference to GDPR" is not the same as "GDPR certified". There is no general GDPR certificate.
  • Absence of SOC 2 or ISO 27001 is not disqualifying for every use case, but it must be disclosed, and you should decide whether your own policy requires it.
  • A national privacy mark such as Japan's PrivacyMark is a specific, checkable credential.

A vendor who tells you plainly what they do not hold is usually more trustworthy than one who says "enterprise-grade security" and nothing else.

A simple scoring approach

You do not need a complicated spreadsheet. Score each of the twelve items as met, partly met or not met, then decide which "not met" items you can live with.

Reassuring answer
Warning sign
Specific feature, where to find it, plan it needs
"Enterprise-grade" with no detail
Clear statement of what is not certified
Vague claims of compliance
Defaults and limits are stated
"Contact sales" for every question
Offers a sandbox to test
Demo only, no trial access
How to read vendor answersIllustrative example

A practical sequence for the evaluation:

  1. Send the twelve questions in writing before any demo.
  2. Compare written answers against what you see in a trial account.
  3. Check the pricing page for which security features sit behind a higher plan.
  4. Run a test with fake customer data, including a hostile prompt.
  5. Record the open gaps and the compensating controls you will apply.

Compensating controls you control

Even a strong vendor does not remove your own responsibilities. Whatever you choose:

  • Turn on two-factor sign-in for every user and keep the owner role to one or two people.
  • Keep sensitive documents out of the knowledge base unless the bot truly needs them.
  • Tell customers in the welcome message not to share card numbers or passwords.
  • Review the activity log monthly and remove accounts of people who have left.
  • Set a retention window that matches your own privacy policy.

Doing this in AI SNS CHAT

Here is how AI SNS CHAT lines up against the checklist, so you can verify each point in a trial account. Details of the security practices are on the /security page, and plans are on /pricing.

Access. Users can set up TOTP two-factor sign-in, and administrators can require it. Ten failed logins lock an account for 30 minutes by default. Roles are Owner, Co-owner, Member, Support and Viewer, with per-member permission overrides. Custom roles and SSO are listed as Corporate-only and hidden, and an SSO/SAML implementation was not found, so do not plan around it. Sign-in options are email, Google and phone SMS.

Audit. Under Workspace settings, the Activity log records every create, edit, delete or upload with name, email, feature and time, and each data source shows who added it and who last edited it.

Data handling. Secrets are encrypted with AES-256-GCM and passwords are hashed with bcrypt. Secrets are redacted from logs, and knowledge-base passages shown to strangers are scrubbed of personal data. A conversation retention window exists as an opt-in limit; the default is to keep conversations forever. Deleted agents go to Trash for 30 days, and nightly downloadable backups keep the last seven.

Models. Text providers are Anthropic (Claude), OpenAI, Google (Gemini) and OpenRouter. Failover goes from the agent's own provider to Anthropic, OpenAI, Google, then OpenRouter, and an administrator setting can keep every reply on one vendor. Advanced models and model fallback need Hobby or higher.

Channels and API. Webhook signature checks cover LINE, Meta (WhatsApp, Messenger, Instagram), Twilio and Discord. The REST API uses HMAC-signed webhooks with a delivery log. API keys live under Workspace settings, are shown once, can be revoked and need Hobby or higher. Published limits include 30 public chat requests per minute per IP. See /connections for the channel list.

Abuse controls. Each agent has prompt-injection and abuse filters plus per-device rate limits. Use Playground to attack your own agent before launch.

Certification wording. AI SNS CHAT holds PrivacyMark No. 20002676. Its practices are designed with reference to GDPR and APPI. The site states explicitly that it is not SOC 2, ISO 27001 or HIPAA certified. Customers are also asked to sign a My Number (特定個人情報) memorandum before creating an agent.

Frequently asked questions

Do I need SOC 2 or ISO 27001 from every AI chat vendor?

Not necessarily. It depends on your own policy, your customers' contracts and the data the bot will handle. If you require them, ask early and treat a plain "we do not hold this" as a valid answer to plan around.

Can I keep all replies on a single AI model provider?

In AI SNS CHAT, the default failover moves across several providers if the first is unavailable, and an administrator setting can keep every reply on one vendor. Confirm which mode your account uses if provider count matters to your review.

Does AI SNS CHAT support SSO?

SSO is gated to the Corporate plan and hidden, and no SAML implementation was found in the product. Use mandatory two-factor sign-in and role restrictions instead.

How long are chat conversations stored?

By default they are kept until you remove them. A retention window can be set as an opt-in limit, so decide what period fits your privacy policy and ask your administrator to apply it.

How do I test whether the bot resists manipulation?

Use Playground to try instruction-override prompts and off-topic requests against your agent, with the per-agent injection and abuse filters switched on. Repeat the test after every persona or knowledge change.

#security#vendor checklist#AI chat#compliance#procurement

Put this into practice with AI SNS CHAT

Build an AI agent trained on your own content and connect it to your website, LINE, WhatsApp and more. Free to start, no credit card required.