A voice agent sits in one of the most sensitive places in a business: the conversation between you and your customers. Callers share names, phone numbers, email addresses, account details, appointment reasons and sometimes much more. The agent connects to your calendar, your email and your documents. Before any enterprise puts an agent in front of customers, its security, IT and compliance teams will rightly want to know how all of that is protected.
This article is a practical checklist. It sets out the questions we think every organisation should ask any voice AI vendor, grouped into access, data, integrations, deployment and accountability. For each group, we explain how Cypher LM answers today, and we are explicit about the gaps we are still closing. A checklist that only lists strengths is a brochure; a useful one tells you where to look harder.
It is written for IT and security leads, procurement teams and operations managers evaluating voice AI for a larger organisation, and for smaller teams who want to get the basics right from the start.
Why voice AI needs its own security review
Voice AI combines several kinds of risk that are usually reviewed separately. It is a customer-facing channel, like a website. It processes personal data, like a CRM. It connects to productivity tools, like a calendar app. And it records and transcribes conversations, which brings its own retention and access questions.
Each of these is familiar on its own. Together, they mean a single vendor relationship touches your customer data, your staff accounts and your Google Workspace. A gap in any one area can expose the others. That is why a voice AI review should cover the whole path: who can sign in, what they can see, where call data lives, what the vendor can do in your connected accounts, and where the agent can be embedded.
The other reason is novelty. Many teams have years of experience reviewing email and CRM vendors and very little experience with AI agents. The questions below are designed to translate familiar security principles into the specific shape of a voice agent platform.
Access: who can sign in, and what can they do?
The first questions are about people. Every account on the platform is a way into your customer conversations.
Questions to ask any vendor:
- Can we give different people different levels of access?
- Can we restrict who sees payment information, exports data or changes agent behaviour?
- Can a user sign out of every device at once if a laptop is lost?
- How are passwords stored, and how are repeated failed sign-ins handled?
- Is two-step verification available?
How Cypher LM answers. Teams are managed in Settings › Team & roles. There are four roles. The Owner controls everything, including the plan, billing and deleting the account. A Manager can build and edit agents, listen to calls and export data. Front desk users see today's calls and their own follow-ups, can view agents, make test calls and mark meeting outcomes, with no access to money anywhere. Sales rep users work on their own follow-ups with limited ability to change agents. Permissions are checked on the server for each action, not just hidden in the interface.
In Settings › Security, any user can sign out of all devices at once, which is the right response to a lost laptop or a shared login. Passwords are stored as salted hashes, never in plain text, and repeated failed sign-ins lock the account for a short period.
The gap. Two-step verification is not available yet; it is on our roadmap and shown as coming soon in the dashboard. Until it ships, we recommend strong, unique passwords or signing in with Google, so your organisation's own Google account policies, including its two-step rules, apply.
Data: where call content lives and who sees it
Calls produce recordings, transcripts, summaries and captured details. These are the most sensitive artefacts a voice agent creates.
Questions to ask any vendor:
- Who can see the names, numbers and emails captured from calls?
- How long are recordings and transcripts kept, and can we delete them?
- Can we export our data in a usable format?
- How is our account and all its data deleted if we leave?
- Is our data used for anything other than serving our agents?
How Cypher LM answers. Caller details such as names, phone numbers and emails are picked up from what callers say and are visible only to your account and the teammates you invite. Each organisation's data is isolated from every other organisation's.
Recordings and transcripts are kept until you delete them. To delete one, open the call and choose to delete its recording and transcript; this removes the recording, transcript, summary and captured details from Cypher LM, while the call still counts towards your totals and minutes. In Settings › Privacy & data you can export your calls as a spreadsheet (CSV) with caller details, duration, mood, language and a short summary. Exports are protected against spreadsheet formula injection, so a malicious caller cannot plant a formula that runs when someone opens the file.
Your business data, uploaded files and call content are used to run your agents and are not shared with third parties for their own purposes.
The gap. Account deletion is not self-serve yet. You request it by email from the dashboard, and our team completes it within seven days. If your policies require an automated deletion workflow, factor that into your review.
Integrations: what the vendor can do in your other systems
Voice agents are most useful when they connect to calendars, email and documents. Those connections are also where a vendor gains the most power over your accounts.
Questions to ask any vendor:
- Which permissions does each integration request, and are they the minimum needed?
- Can we connect one service without granting all of them?
- How are the stored tokens and keys protected?
- Can we disconnect at any time, and what happens when we do?
- Can a connection started by someone else be attached to our account?
How Cypher LM answers. Google services are connected one at a time in Settings › Integrations: Calendar, Meet, Gmail, Docs, Sheets and Drive each ask for their own permission, and Cypher LM only uses the permissions you grant. When Cypher LM writes to your Google Drive, it uses Google's file-level permission, which only lets it create and update files that Cypher LM itself created: the call-log spreadsheet, the call-notes document and the meeting notes documents. Reading your own documents, for example to import them as knowledge, needs a separate permission that you choose to grant. Importing from Drive is limited to the account owner.
Stored Google tokens and Stripe keys are encrypted at rest, and Stripe keys are checked with Stripe before they are saved. You can disconnect any integration at any time; Cypher LM then treats that service as not connected.
Deployment: where the agent can run
An agent that can be embedded anywhere can be misused anywhere. Deployment controls decide where your agent is allowed to appear.
Questions to ask any vendor:
- Can we restrict which websites can embed our agent?
- Are public pages, such as booking or payment pages, protected from abuse?
- How are webhooks and callbacks authenticated?
- What protects the platform from automated abuse and request floods?
How Cypher LM answers. The website widget only works on the website addresses you list in Deploy › Website widget (subdomains included). If someone copies your embed code onto another site, the widget does not appear there. Public booking and payment pages use long, unguessable tokens, and public forms have shared rate limits per visitor, per customer and per business, so they cannot be used to send floods of email from your account.
When the Cypher M meeting bot reports back from a Google Meet, each callback is signed for that specific session; anything unsigned is rejected. Webhooks from our voice provider are verified the same way. The platform is rate-limited at the edge, and outbound requests made on behalf of customers are checked against private and internal network addresses.
The developer API and signed outgoing webhooks are in beta and not yet open to accounts. When they open, API keys will be shown once and stored only as hashes, each key will be rate-limited, and every webhook delivery will carry a signature you can verify.
The agent itself: what it says and how you control it
Security reviews often stop at infrastructure, but with a voice agent the content of the conversation is part of the risk. An agent that invents a refund policy or discusses one customer's booking with another caller is a problem, however well the servers are protected.
Questions to ask any vendor:
- Where do the agent's answers come from, and can we limit it to our own material?
- Can we tell the agent which topics it must never discuss?
- Can we review what the agent actually said on every call?
- Can we test the agent before customers hear it?
How Cypher LM answers. Every agent is grounded in the business details you enter and the knowledge files you upload or import, so its answers come from your own documents rather than general knowledge. In the agent's behaviour settings you list its responsibilities, the topics it must avoid, what it says when it does not know, and when a caller should be passed to a person. Those instructions are managed by your team and are never shown to callers.
Every call is logged in Calls with its transcript and, with Cypher N, a recap: a summary, key details, action items and a drafted follow-up. That gives your team a complete record of what the agent said, which is the best way to spot an answer that needs correcting. Test calls let you try the agent yourself before it goes live, and they are kept out of your analytics and exports so they do not distort your numbers.
For an enterprise rollout, we recommend a short review routine during the first month: a named person reads a sample of recaps each week, checks answers against policy, and updates the knowledge files or restrictions where needed. Treat the agent's instructions like any other controlled document, with a clear owner and a record of changes.
Accountability: certifications, reviews and honesty
The last group of questions is about trust in the vendor itself.
Questions to ask any vendor:
- Which independent certifications do you hold?
- When did you last review your own security, and what did you fix?
- How do you handle security reports?
- What will you tell us if something goes wrong?
How Cypher LM answers. We do not currently hold third-party security certifications, and we do not claim to. What we can tell you is that the platform, API and dashboard alike, is reviewed for authentication, tenant isolation, role permissions, injection and information leaks, that security fixes are covered by automated tests, and that dependencies are checked for known vulnerabilities.
If your organisation requires a specific certification before go-live, ask us directly at sales@cypherlm.com. We would rather tell you what we have and what we are working on than have you discover a gap later. Our security page summarises our current posture.
The checklist at a glance
| Area | Question | Cypher LM today |
|---|---|---|
| Access | Role-based permissions | Owner, Manager, Front desk, Sales rep, checked on the server |
| Access | Sign out everywhere | Yes, in Settings › Security |
| Access | Two-step verification | Coming soon; use Google sign-in meanwhile |
| Data | Who sees caller details | Only your account and invited teammates |
| Data | Retention and deletion | Kept until you delete; delete per call |
| Data | Export | CSV export in Settings › Privacy & data |
| Data | Account deletion | By email request, completed within 7 days |
| Integrations | Least privilege | Per-service Google permissions; Drive writes only to files Cypher LM created |
| Integrations | Stored credentials | Google tokens and Stripe keys encrypted |
| Deployment | Embed control | Widget works only on your listed websites |
| Deployment | Callbacks | Signed per session; unsigned requests rejected |
| Accountability | Certifications | None claimed today |
How this works in Cypher LM
If you are rolling Cypher LM out across a larger organisation, this order works well:
- Decide who owns the account. One Owner, ideally a shared operations or IT mailbox rather than an individual, so access does not depend on one person.
- Invite the team with the least access they need. Use Manager for people who build and edit agents, Front desk for people who handle calls day to day, and Sales rep for people who own follow-ups.
- Connect only the Google services you use. Start with Calendar and Meet for bookings; add Gmail, Docs, Sheets or Drive only when you need them.
- Lock down the widget. List only your production websites in Deploy › Website widget.
- Set your retention routine. Decide how long you need recordings and transcripts, and delete calls you no longer need on that schedule.
- Review regularly. Check the team list and connected integrations each quarter, and remove anyone who has left.
Frequently asked questions
Does Cypher LM hold a security certification?
Not today. We do not claim certifications we have not earned. Contact sales@cypherlm.com if your organisation has specific requirements, and we will tell you exactly where we stand.
Can we restrict who sees payment information?
Yes. Front desk users have no access to payment amounts or actions. Only roles with the right permissions can see payments, mark them as paid or issue refunds.
What happens to our Google data if we disconnect?
Cypher LM stops using that service immediately. Files Cypher LM created in your Drive, such as the call-log sheet and notes documents, stay in your Drive and remain yours.
Can someone embed our agent on their own website?
No. The widget only appears on the website addresses you list. Copying the embed code to another domain does not make the widget work there.
Is two-step verification available?
Not yet; it is coming soon. In the meantime, signing in with Google lets your organisation's Google account policies apply.
Review with confidence
Security reviews go faster when both sides are specific. Use this checklist with any voice AI vendor, and expect clear answers, including about gaps. We have tried to give you ours.
To start a review, book a demo with our team or read more on the security page and our enterprise solutions page. Plans, including enterprise options, are on the pricing page.
