Security and
accountability.
See how InVice protects client information and keeps people accountable for agent actions.
Received a message from an InVice agent?
The agent sends from a person’s connected mailbox on behalf of their business. Reply to that message to reach the sender. The business sets its permissions, and actions that require a decision wait for approval.
Trust is the
first feature
A system of action over the systems of record has to earn the right to act in your name. InVice assumes that model output and connected content can be wrong or adversarial, then keeps authority in code and explicit user decisions. Encryption, key handling, action authorization, and audit are built before the features that depend on them. The short version is below; the full brief is published in its entirety.
You set the limits.
The record stays visible.
An agent can act only within the permissions you set. Work that requires your decision waits for review, and actions leave a record you can inspect.
Read the security briefChoose the context, tools and actions available to an agent.
Approval shows the exact action and destination before it runs.
Audit records show what happened without storing the message contents.
Security, in detail.
Open any control for its scope and limits. The full security brief is published on this site.
Prompt injectionUntrusted text cannot grant authority
Email, documents, web pages, connector output, model output, and generated summaries are treated as untrusted content. Action risk and data classes are code-defined; proposed destinations are payload-bound and checked again at execution.
Exact approvalsPayload-bound and single-use
When policy requires review, the interface shows the action, destination, and data classes involved. Approval is bound to that exact payload, expires, and is consumed atomically; changing the recipient or content invalidates it.
Outbound actionsOne enforcement gate
Automated external mutations pass one gate immediately before the provider call. It verifies workspace and mission scope, configured autonomy or exact approval, destination, automation pause, and dry-run state.
Outbound emailYou are the sender
Client emails send through your own connected Gmail or Outlook account, and by default nothing sends until you approve the draft.
Connected toolsFail closed by default
Dynamic MCP and REST tools require exact review before connector credentials are decrypted. Tenant-controlled destinations pass redirect-aware SSRF validation, and model calls never silently fall over to a different provider.
Encryption at restIn your private vault
Documents are encrypted at rest. Decrypted only when an agent needs them, and only inside the permission envelope of the active mission.
Encryption in transitTLS and HSTS
Traffic to and from InVice is encrypted in transit. HSTS prevents browsers from falling back to plaintext connections.
API keysNever stored plaintext
Your model API key is AES-256 encrypted at rest, decrypted only into isolated session memory, and discarded immediately after the call. Never logged.
AuthSupabase Auth, MFA-ready
Sessions verified via JWT on every request. We use getUser() not getSession() on the server. MFA available; SSO and SAML on Enterprise.
Audit logsMetadata-only by policy
Every agent action writes an audit row. Payloads are scrubbed of PII at write. We record what happened, not the contents.
Where your data is storedConfigured regions and processing exceptions
Your database and uploaded files are stored in a Canadian region (Montréal). Your own model provider processes prompts under your API key and may run elsewhere. Vault embeddings currently process in the US (OpenAI) and are moving to a Canadian region; we will name the provider when it ships. Learning inference also uses Anthropic in the United States; see our Privacy Policy for the data and safeguards involved. We recommend zero-retention model providers. Enterprise can request region-locked deployments.
TrainingInVice does not train on your data
InVice does not use your documents, missions, or audit logs to train models. A model provider you connect processes selected prompts under its own contract and settings.
RetentionYours to control
Export window on cancellation, then permanent deletion confirmed in writing. Request deletion at any time.
Read the
full security brief
The long version, published on the site instead of sent on request: threat model, encryption details, key handling and rotation, audit format, retention controls, sub-processor list, and where we stand on SOC 2. Where the product and the promise differ, the brief says so.