Security

A description of the controls the platform actually implements, so they can be verified rather than assumed.

Authentication

Passwords are hashed with scrypt and a per-user salt.
Login compares against a dummy hash when the address is unknown, so a failed login takes the same time whether or not the account exists.
Sessions are opaque random tokens in an httpOnly, SameSite=Lax cookie (Secure in production). Only a SHA-256 hash is stored server-side.
Changing or resetting a password revokes every session on every device.
Verification and reset tokens are single-use, hashed at rest, and expire (24 hours and 1 hour respectively).
Google sign-in verifies a CSRF state parameter, and refuses to link an unverified Google address to an existing account.

Authorization

Role-based access control with a permission table, not scattered role comparisons. Roles: USER, AGENT_BUILDER, SUPPORT, MANAGER, ADMIN, SUPER_ADMIN.
Every query that reads customer data is scoped by user id, so an identifier from another account resolves to a 404 rather than a leak.
Admin routes check the specific permission they need, independently of the layout that renders the page.
Nobody can change their own role or status, or grant a role above their own rank.

Request protection

CSRF: cookie-authenticated mutations require an Origin match and a double-submit token.
Rate limiting on login, registration, password reset, agent runs, public forms and webhooks.
Input validation with Zod at every API boundary. Invalid input returns field-level errors, never a stack trace.
Database access goes through Prisma with parameterised queries; there is no string-built SQL.
Outbound requests from agents and workflow nodes reject loopback and private network ranges, blocking SSRF.

Secrets

Integration credentials and OAuth tokens are encrypted with AES-256-GCM before storage and never sent to a browser.
API keys are stored as SHA-256 hashes. The raw key appears exactly once, in the creation response.
Provider API keys live only in server environment variables. All model calls happen server-side.
Logs pass through a redaction filter that strips credential-shaped fields.

Agent execution

Tools declare a risk tier: READ, WRITE or DESTRUCTIVE. An agent version declares the highest tier it may perform automatically.
A call above that ceiling is refused, recorded as awaiting approval, and reported to the model as not performed.
Runs are bounded by step count, token budget and a wall-clock timeout.
Every step is persisted with its input, output, duration and risk tier.

Payments

PROMIVO never sees or stores card data. Stripe handles it end to end.
Webhook signatures are verified before the payload is read as data.
Every event is persisted under a unique constraint, so a redelivery is recognised and skipped rather than double-applied.
Entitlements are granted only by the webhook handler, never by a client-side success redirect.

Reporting a vulnerability

Reach us through the contact page with steps to reproduce. Please give us a reasonable window to fix an issue before disclosing it publicly.

This is a plain-language summary of how the platform behaves, written to match what the software actually does. It is not a substitute for legal advice, and should be reviewed by counsel before commercial launch.

Authentication

Passwords are hashed with scrypt and a per-user salt.

Login compares against a dummy hash when the address is unknown, so a failed login takes the same time whether or not the account exists.

Sessions are opaque random tokens in an httpOnly, SameSite=Lax cookie (Secure in production). Only a SHA-256 hash is stored server-side.

Changing or resetting a password revokes every session on every device.

Verification and reset tokens are single-use, hashed at rest, and expire (24 hours and 1 hour respectively).

Google sign-in verifies a CSRF state parameter, and refuses to link an unverified Google address to an existing account.

Authorization

Role-based access control with a permission table, not scattered role comparisons. Roles: USER, AGENT_BUILDER, SUPPORT, MANAGER, ADMIN, SUPER_ADMIN.

Every query that reads customer data is scoped by user id, so an identifier from another account resolves to a 404 rather than a leak.

Admin routes check the specific permission they need, independently of the layout that renders the page.

Nobody can change their own role or status, or grant a role above their own rank.

Request protection

CSRF: cookie-authenticated mutations require an Origin match and a double-submit token.

Rate limiting on login, registration, password reset, agent runs, public forms and webhooks.

Input validation with Zod at every API boundary. Invalid input returns field-level errors, never a stack trace.

Database access goes through Prisma with parameterised queries; there is no string-built SQL.

Outbound requests from agents and workflow nodes reject loopback and private network ranges, blocking SSRF.

Secrets

Integration credentials and OAuth tokens are encrypted with AES-256-GCM before storage and never sent to a browser.

API keys are stored as SHA-256 hashes. The raw key appears exactly once, in the creation response.

Provider API keys live only in server environment variables. All model calls happen server-side.

Logs pass through a redaction filter that strips credential-shaped fields.

Agent execution

Tools declare a risk tier: READ, WRITE or DESTRUCTIVE. An agent version declares the highest tier it may perform automatically.

A call above that ceiling is refused, recorded as awaiting approval, and reported to the model as not performed.

Runs are bounded by step count, token budget and a wall-clock timeout.

Every step is persisted with its input, output, duration and risk tier.

Payments

PROMIVO never sees or stores card data. Stripe handles it end to end.

Webhook signatures are verified before the payload is read as data.

Every event is persisted under a unique constraint, so a redelivery is recognised and skipped rather than double-applied.

Entitlements are granted only by the webhook handler, never by a client-side success redirect.