Skip to content
TalkDemo
Security

Built so a security review is short

Demo tenants only, credentials the AI never sees, guardrails on every click, and data kept in India for Indian customers.

Last updated

By design

How your product and your prospects are protected

TalkDemo is pre-launch, in private pilots. Every control below is built with the product and marked Planned until it ships with its test, then it moves to In place. Each is mapped to ISO/IEC 27001 Annex A and SOC 2 in an internal register.

Your demo tenant and the browser

The agent drives a copy of your product made for demos, inside limits you set, and nothing else.

  • Demo tenants only

    Planned

    Live product mode connects only to a dedicated demo tenant with sample data. Our terms forbid connecting a production account or real customer data.

  • One isolated browser per session

    Planned

    Each call gets its own cloud browser, logged into your demo tenant. It is destroyed when the call ends, and demo data can be reset through your tenant's API or a snapshot.

  • URL allowlist

    Planned

    The browser can only open the hosts you list. A link to anywhere else is refused before the page loads.

  • Blocked actions and read-only mode

    Planned

    Delete, billing, settings, invites and exports are blocked by default, and you can block any button, selector or text. Forms can be read only.

  • Checked before every action

    Planned

    The browser service checks each click and keystroke against your guardrails, whatever the AI model asked for. A refused action is logged and the agent explains it can't do that in a demo.

Credentials and secrets

The AI model never sees the password to your demo tenant.

  • Sealed demo credentials

    Planned

    The demo login is encrypted with AES-256-GCM in its own column. Only Admins of your workspace can set or replace it, and nobody can read it back.

  • Opened only by the browser service

    Planned

    The browser service opens the credentials to fill the login form. They are never part of a prompt, a log line, an API response or a page sent to a browser.

  • Sealed integration tokens

    Planned

    Tokens for your CRM, Slack and calendar are sealed the same way and used only by the job that calls that service.

  • Secrets outside the code

    Planned

    Keys live in the hosting provider's secret store, separate for each environment, with a written rotation procedure.

Prospect data

Prospects join your demo on your behalf, so their data is kept to what the demo needs.

  • Sealed contact details

    Planned

    Prospect email addresses and phone numbers are encrypted, with a keyed hash beside them for de-duplication. Lists show them masked, and Viewers can't reveal them.

  • No raw IP addresses

    Planned

    A session keeps a hash of the prospect's IP address and their country, never the address itself.

  • Notice and consent before recording

    Planned

    The call page tells the prospect the call is with an AI agent and asks before recording starts. Recording is on by default and you can turn it off per agent.

  • No camera, ever

    Planned

    The call asks only for the microphone, and typing works if the prospect says no. The shared screen is the agent's browser showing your demo tenant, never the prospect's screen.

  • Deleted on schedule

    Planned

    Recordings, transcripts and prospect details are deleted by a daily job after 90 days by default, or sooner if you set a shorter period.

Separation between customers

One customer can never reach another customer's agents, sessions or leads, even through a bug in a screen.

  • Enforced by the database

    Planned

    Every agent, flow, session, transcript and lead row carries its workspace, and composite keys make the database refuse a link between records of different workspaces.

  • One permission check everywhere

    Planned

    The same check runs for pages, APIs, knowledge search, background jobs, webhooks and exports.

  • Unknown records answer "not found"

    Planned

    Asking for another workspace's record answers "not found", without saying whether it exists, and the attempt is recorded as a security event.

AI processing

The model gets what the conversation needs and nothing more.

  • Only what the task needs

    Planned

    The model receives the conversation, relevant passages from your knowledge sources and short summaries of the screen. Never demo credentials, and never another workspace's data.

  • No training on your data

    Planned

    Customer and prospect data is never used to train models. We use zero-retention settings with model providers where they offer them.

  • Content treated as data

    Planned

    Help docs, web pages and what appears on screen are handled as data, never as instructions, so a page can't tell the agent to change its rules.

  • An honest agent

    Planned

    The agent says it is an AI when asked, answers from your knowledge, and says it will check with your team when it isn't sure instead of guessing.

Accounts and staff access

Few people can see data, and each sees only what their role needs.

  • Roles with least privilege

    Planned

    Admins manage the team, billing, credentials and retention. Members build agents. Viewers read sessions and leads but can't change agents or reveal contact details.

  • Two-factor sign-in for staff

    Planned

    TalkDemo staff must use two-factor sign-in, and staff roles (owner, admin, support) each get only the access their job needs.

  • Support access only with your permission

    Planned

    Staff can open your workspace only after one of your Admins grants time-limited, read-only access. Every page they open is recorded.

  • One-step account lock

    Planned

    Staff can lock an account and end all its sessions at once, with a reason recorded in the audit trail.

Audit trail and monitoring

Every agent action leaves a record on the session timeline, with its time.

  • Separate logs database

    Planned

    The audit trail, security events and AI call logs live in their own database, apart from customer data.

  • Every agent step recorded

    Planned

    Each browser action, tool call and model step is logged with the session and linked to its place in the recording.

  • Audit trail of people

    Planned

    Actions by your team and by TalkDemo staff are recorded with who, what and when.

  • Nothing sensitive in logs

    Planned

    Logs never contain passwords, demo credentials, tokens, full contact details or raw IP addresses. AI call logs keep counts and costs, not text.

Public entry points

Demo links, the widget and public forms are checked before anything reaches the data.

  • Bot checks and rate limits

    Planned

    Public forms use Cloudflare Turnstile to tell people from bots, and every public endpoint is rate limited with strict input validation.

  • Widget allowed origins

    Planned

    The widget starts a call only from the website addresses you list, checked on the server, and the call page can be framed only by those sites.

  • Signed webhooks

    Planned

    Events from payment and messaging providers are accepted only with a valid signature. Webhooks we send you are signed with HMAC so you can check them.

  • Uploads scanned

    Planned

    Decks, screenshots and knowledge files are scanned for malware before they are shown in a call or indexed.

  • No card data at TalkDemo

    Planned

    Payments run on Razorpay and Stripe hosted checkout. Card and UPI details stay with them.

Retention

How long data is kept

Recordings, transcripts and prospect details are kept for 90 days by default. You can shorten that per agent, and a daily job deletes data when its period ends.

Recordings, transcripts and prospect details
90 days, or shorter if you set it
Demo credentials
Until you remove them
Knowledge sources
Until you delete them
AI call logs (counts and costs, no text)
90 days
Audit trail and security events
1 year
Consent records
Kept as proof after deletion, with personal details removed
Data requests

Prospects can ask for a copy, a correction or deletion

Incident response

If something goes wrong

The plan for a security incident, from the first alert to what we change afterwards.

  1. Detect

    Serious security events, refused browser actions in bulk, spikes in failed sign-ins, quarantined files and reported demo links alert TalkDemo staff.

  2. Contain

    Staff can lock an account and end its sessions in one step, pause an agent so its demo link stops taking calls, end live browser sessions, or rotate a key.

  3. Tell you

    If a breach affects your data or your prospects' data, we tell your Admins without undue delay, and within 72 hours at the latest, with what we know, so you can meet your own duties.

  4. Notify and learn

    We help you notify the Data Protection Board of India and affected prospects under the DPDP Act, and your supervisory authority under the GDPR or UK GDPR. Then we record what happened and what we changed.

Compliance roadmap

Aligned with the frameworks, not certified

TalkDemo holds no certifications and shows no badges. This is a roadmap, not a certification: we keep a control register mapped to SOC 2, ISO/IEC 27001, India's DPDP Act and the GDPR, and this is where each one stands.

FrameworkWhere we areNext step
SOC 2Not audited. Controls are mapped to the Trust Services Criteria in an internal register and built with the product.A Type I audit when the first enterprise customer needs it, then Type II.
ISO/IEC 27001:2022Not certified. The same controls are mapped to Annex A.Set up the management system (scope, policies, risk treatment) alongside the SOC 2 work.
India's DPDP Act 2023Designed for customers as data fiduciaries: notice and consent before recording, retention limits and data in India.Finish the legal review of the privacy policy and name a grievance officer.
GDPR and UK GDPRDesigned for the processor role: recording consent, access and erasure tools, customer-set retention and a public sub-processor list.Finalise the data processing agreement with EU standard contractual clauses and the UK transfer agreement, and records of processing.

Also planned

The operational work around the product controls: testing, backups, reviews and policies. None of it is in place today.

  • Planned: Encryption in transit with TLS on every connection, and encryption at rest for the database and file storage
  • Planned: Automated tests on every code change, including cross-workspace and guardrail tests, blocking merges that fail
  • Planned: Dependency and secret scanning on every code change
  • Planned: Prompt-injection tests on docs, web pages and on-screen text before any model or prompt change goes live
  • Planned: Point-in-time database restore, nightly encrypted backups and a regular restore drill
  • Planned: An external penetration test before the first paid customers after the pilots
  • Planned: Quarterly review of staff access and a staff offboarding checklist
  • Planned: Written security policies, a risk register and a breach notification runbook
  • Planned: Signed data processing agreements and security reports collected from every provider
Report a vulnerability

Found a security problem? Tell us.

Rules for testing

We will not pursue good-faith research that follows these rules.

FAQ

Security questions

No. TalkDemo is pre-launch and has not been audited or certified. Our controls are mapped to SOC 2 and ISO/IEC 27001 Annex A in an internal register and built with the product. We plan a SOC 2 Type I audit when the first enterprise customer needs it, and will say so here only when an audit is complete.

No. The demo login is sealed with AES-256-GCM and opened only by the browser service, which types it into your login form. It is never part of a prompt, a log, an API response or anything sent to a browser. The model only knows it is logged in as the demo user.

Every browser action is checked against your guardrails before it runs: a URL allowlist, blocked actions such as delete, billing, settings, invites and exports, blocked buttons or text you choose, and an optional read-only mode. The agent only ever runs in your demo tenant, never production.

For customers in India: in India, with the database and file storage in AWS Mumbai (ap-south-1) and Indian-language speech on Sarvam, an India-based provider. Calls in world languages, such as English (US) or Spanish, use speech providers in the United States. Calls are routed to the nearest region. The AI model runs at Anthropic in the United States, with only what the task needs. The sub-processors page lists every provider.

Only if one of your Admins grants support access for a limited time. Support then sees your workspace read only, and every page they open is recorded. Staff must use two-factor sign-in.

Security, privacy or procurement questions? Email hello@talkdemo.com.