Security and trust

Agents with real access need real controls.

An agent that can query your warehouse, browse your suppliers and send email is only useful if you can trust where it stops. This is how the platform keeps it inside the lines, and what you can check for yourself.

Last reviewed October 2026

Built in, not bolted onenforced

Keys encrypted at rest

authenticated encryption

Values never shown to the model

used through a broker

Code in fresh sandboxes

no credentials inside

Isolation at the database

row-level access control

Every step recorded

chat, runs, usage

Defence in depth

Six layers between a request and your data.

No single control is trusted to hold alone. A request has to pass identity, organization scope, the agent's own permissions and a vault before anything runs, and it leaves a record whatever the outcome.

We design for the day a model is fooled by a hostile web page or message. So limits do not depend on the model behaving: the key is not in reach, the sites are listed, sending is capped and the sandbox is empty.

01

Identity

Sign-in through an audited auth provider. We never store a password in plaintext.

02

Organization isolation

Every row is checked against ownership and connection at the database, not just hidden in the interface.

03

Agent permissions

An agent reaches only the tools, keys and sites attached to it. Sharing is switched on per tool.

04

Secrets vault and broker

Keys are encrypted at rest and used on the agent's behalf. The model never sees a value.

05

Isolated execution

Code runs in a fresh, short-lived sandbox with no database access and no credentials of its own.

06

Record and control

Every step, call and cost is recorded. Pause, cap or turn off any agent, tool or person.

Keys and secrets

The model never holds the key.

It is not enough to hide a key from the model's output. Code that already holds a value can be talked into leaking it a character at a time. So the sandbox never receives the value at all.

Encrypted at rest

Admin keys, per-person keys and connected-account tokens use authenticated encryption and are decrypted only in the backend, at the moment of use. No page or API returns a value.

Short-lived and scoped

Each sandbox call gets a one-use token that expires in minutes. It can ask for the named keys of that call, as that person, and nothing else.

Fail-closed by default

A key can reach only the hosts you list for it. Keys that need signed requests are kept out of agent-written code and used only inside tools you write.

The same rule runs through the other doors. MCP credentials are decrypted only inside the backend, and the model never sees the server address. Warehouse and storage connections run a fixed list of curated actions, not arbitrary signed requests. Website logins are typed into a secure form that the assistant cannot read.

Organization and person isolation

Your data is separated where it is stored, not just where it is shown.

Ownership and connection decide what a row can be read or changed by, enforced by the database itself. A bug in a page cannot turn into a leak, because the page never had the access in the first place.

  • Organizations, agents, chats and messages are scoped to their owner
  • A connected person gets their own chats, files and keys, hidden from every other person on the agent
  • Pausing a person stops access without altering their history
  • Deleting an agent removes its files and records, including every connected person's

Execution

Code runs somewhere that has nothing to steal.

Anything an agent writes and runs, and anything in a tool, executes in a separate isolated service, not in the system that holds your conversations and credentials.

  • A fresh, short-lived sandbox for every call, with no database access and no credentials of its own
  • Files go in explicitly for that one call, and only designated output files come back
  • Hard limits on run time, memory and concurrency so one job cannot starve others
  • Browser sessions each get their own isolated environment, a per-person cap, and are closed when idle
Before an agent sends an emailchecked
Row accessowner or connected only
Tool allowedsend_email · for you
Recipienton the allowlist
Daily limit14 of 50 used
Balancein good standing

Your data

What is stored, who can read it, and when it goes.

Retention is written into the system, and several clean-ups run in the database itself so they hold even if another service is down.

DataWho can read itKeptRemoved when
Chats and messagesThe person they belong to. Other connected people cannot see them. The organization that owns the agent can administer its agents' data.Until you delete themRemoving an agent from your list deletes your chats and workspace for it.
Keys and secretsNo one. Values can be written, never read back.Until you delete themDeleting the key or the agent.
Workspace filesThe person they belong to. Other connected people cannot see them.Until you delete themLeaving an agent removes your files; deleting an agent removes everyone's.
WhatsApp messages and filesThe agent, only for chats you opened, and only for the number linked now.30 days, hard capUnlinking deletes everything at once. A database job also removes anything past 30 days every hour, even if a service is down.
Connection historyThe owner and organization admins.30 daysExpired automatically.
Temporary sandbox accessThe backend only.MinutesSwept every 15 minutes, and removed when the call ends.
Settled invitationsThe two people involved.5 days after settledRemoved automatically.
PaymentsCard, UPI and bank details are held by the payment processor, never by us. We keep the amount, time and a reference.As records requirePer the Privacy Policy.

The infrastructure providers that process data on our behalf are listed in the Privacy Policy. Text sent to an AI model to produce a reply is processed by the model provider the platform is configured to use.

Controls you hold

Limits you set are limits the agent keeps.

Every one of these is enforced by the platform on each action, and none depends on the model choosing to comply.

Tools are opt-in

Nothing is available to an agent until you switch it on. Each tool also has its own switch for connected people.

Approved sites only

The browser reaches the sites you list. Logins go in through a form the assistant never reads.

Approved recipients

Email goes only to the people you allow, within a daily limit, with delivery status on every message.

Chat-level WhatsApp access

You open each group or contact as read, send or both. The gateway re-checks every send, with a daily cap and duplicate protection.

Per-key host allowlists

A key can be used only against the hosts you name for it. With none named, it cannot be used in a generic request at all.

Balance and plan limits

When a plan lapses or a balance runs out, the agent pauses. It cannot spend what is not there.

Pause anyone, instantly

Pausing a connected person stops their access without touching their history.

A record of everything

Messages, status changes, every task step and every model call are recorded with their cost.

Practices

The ordinary things, done properly.

Sign-in

Handled by a dedicated, audited authentication provider with email and Google sign-in. We do not implement our own authentication or store your password in plaintext.

Connecting other services

Uses that provider's own consent screen with the standard authorization-code flow and PKCE. We never see or ask for the password of a service you connect.

Payments

Processed by Razorpay. A top-up is credited only after a signed confirmation arrives from Razorpay's servers, never on a browser's say-so, so a closed tab cannot create or lose a credit.

Untrusted content

Messages read from WhatsApp are handed to the agent as untrusted text, not as instructions, and a send needs both permission on the chat and passing the daily and duplicate checks.

Accountability

Messages, status changes, task steps and model calls are recorded, so a decision can be traced to the input that caused it.

What we don't show

No certification badges: we would rather not imply any we do not hold. Architecture walkthroughs and security questionnaires are answered directly on request.

Responsible disclosure

Found something? Tell us first.

If you believe you have found a security issue in Parasync Agents, please report it to team@parasync.in instead of disclosing it publicly.

What helps

  • What you found and where, with steps to reproduce it
  • What you think the impact is
  • How to reach you for follow-up

We take every report seriously, acknowledge it promptly and keep you updated while we investigate and fix it.

For security teams

Evaluating us for your organization?

We will walk through the architecture with you and answer your questionnaire directly.