Skip to content

Security and compliance

Bluedeer is built for organisations that must govern what an AI assistant can see, send and change. The control set spans policy enforcement and approval workflows, data loss prevention (DLP) controls for AI actions, PII detection and redaction, data residency, identity and access management, and a tamper-evident audit log that customers can verify independently.

  • Policy enforcement precedes user preferences. Workspace rules block or escalate actions by type, amount, cumulative daily value, record count, risk level and recipient domain, across chats, scheduled tasks, event-triggered tasks and live pages.
  • Approval workflows enforce segregation of duties. An action is routed to a designated approver, the requester can never approve it, unanswered requests expire after 24 hours, and any change to the action voids the approval.
  • The audit log is tamper-evident. Every entry is hash-chained to its predecessor, a checkpoint is signed daily, and each export ships with a verification script that runs offline.
  • Data residency is enforced per workspace. AI requests are processed only in the selected regions, with no cross-region fallback.
  • PII detection and redaction apply per category before content reaches the AI provider, with placeholders restored in the response.
  • Customer content is never used to train AI models.

Data loss prevention (DLP) controls for AI actions

Bluedeer applies data loss prevention (DLP) controls to the actions its AI assistant takes on a user's behalf: an allowlist of outside addresses and domains, corporate-identity enforcement for connected accounts, explicit consent and routed approval before data is sent, shared or changed, PII hiding before content reaches the AI, an AI-disclosure line (“Sent by Bluedeer”) on emails the assistant sends on its own, and a tamper-evident audit record of each action. These controls govern AI actions within Bluedeer and complement an organization's DLP platform rather than replace it; they do not inspect file attachments or monitor endpoints, networks or other applications.

How is outbound data exfiltration through the assistant prevented?

Outbound recipients are governed by a domain and address allowlist with an allow, approve or block policy, connected accounts are restricted to corporate identities, and every write action requires explicit consent unless the user has delegated it.

Data loss prevention (DLP) for outbound recipients

Enterprise

Administrators maintain an allowlist of external domains and addresses and set the policy for every other recipient: allow, require approval, or block.

Enforcement covers all email recipients in To, Cc and Bcc (including replies, forwards, drafts and scheduled sends), calendar invitees and file-sharing grantees. Verified domains and their subdomains are always internal. A recipient that cannot be resolved is treated as external. Each allowlist entry is scoped to contact, account connection, or both.

Limits: Slack messages and custom connectors are outside the scope of this control.

How it works in the help centre

Corporate-identity enforcement for connected accounts

Enterprise

Members can connect only corporate accounts to Gmail, Google Workspace apps and Slack; personal accounts are rejected at connection time.

Google accounts are validated against the managing (hosted) domain reported by Google, so a consumer account registered with a corporate address is still rejected. Enabling the control never revokes existing connections automatically: non-compliant accounts are listed for administrators to disconnect.

Limits: Identity cannot be validated for custom connectors or catalogue connectors.

How it works in the help centre

Explicit consent for write actions

Every plan

Every action that sends, shares, modifies or deletes data requires the user's confirmation on an approval card that shows the exact operation, unless the user has delegated it.

Consent is configured per action and per connector tool. Read-only tools default to no prompt. Connecting accounts, managing members, issuing or revoking share links and changing settings are reserved to the user and cannot be performed by the assistant.

How it works in the help centre

AI disclosure on emails sent by Bluedeer

Enterprise

Administrators can require every email that a scheduled task, an event-triggered task or a live page sends on its own to end with the line “Sent by Bluedeer”.

The line reads “Sent by Bluedeer” by default and can be replaced with the workspace's own wording. Senders cannot remove it, and every change to the setting is recorded in the audit log.

Limits: Emails a user sends from chat, including from an email card or an approval card, carry that user's own signature instead.

How it works in the help centre

Least-privilege permissions for scheduled and event-triggered tasks

Depends on the plan

Scheduled and event-triggered tasks run only the actions the owner granted when the task was created; any other action is skipped and reported in the run history.

Content that starts a task, such as an inbound email or a webhook payload, is passed to the AI as untrusted data with an instruction never to follow commands embedded in it. This mitigates prompt injection; the permission grant and workspace policies remain the enforcement boundary.

Limits: Prompt-injection mitigation reduces risk but cannot eliminate it, which is why execution stays gated by permissions and policy.

How it works in the help centre

How are high-risk actions governed, and who approves them?

Workspace policies block or escalate actions before any user-level setting applies, and escalations follow an approval workflow with segregation of duties and a full audit record.

Policy enforcement (workspace rules)

Enterprise

Administrators define policies that block an action for everyone or require approval first, scoped by action type, application, execution context and triggering event.

Conditions include amount above a threshold, cumulative daily value per user (which prevents splitting a payment to stay under a limit), record or recipient count, risk level, unreadable amount, and external recipients. Values that cannot be read are evaluated as the highest-risk case. Policies apply in chats, scheduled tasks, event-triggered tasks and live pages, and to emails and tasks users save themselves. User preferences and task permissions cannot override them.

Limits: Where a custom connector's tool exposes no readable amount, a policy applies to the tool as a whole rather than to individual values.

How it works in the help centre

Approval workflows with segregation of duties

Enterprise

Escalated actions are routed to any administrator, the owner, two administrators, or a named member; the requester can never approve their own request.

Approvers review the exact operation: target, amount, recipients, message body and tool arguments. An approved action executes under the requester's credentials. A rejection may carry a reason of up to 500 characters and is recorded in the task run as “Declined by reviewer”. Requests expire after 24 hours with no action taken, and any change to the action after submission voids the approval.

How it works in the help centre

Can the audit trail be independently verified?

Yes. The audit log is hash-chained and signed daily, and every export includes the material and the script needed to verify it offline, without relying on Bluedeer.

Audit log

Enterprise

A complete record of administrative and user activity in a shared workspace, with actor, target, and before and after values for every change.

Coverage includes membership and roles, seats and billing, sharing and share links, connector configuration, actions the assistant took in connected applications and whether each was approved, tasks, policies and approvals, authentication and single sign-on changes, and support sessions. Entries respect item-level visibility. The log has no retention cut-off, and recording runs on every plan, so full history is available when a workspace moves to Enterprise.

Limits: Read access to an item is not logged as a change, and the content of an action is never stored in the log.

How it works in the help centre

Hash chaining and signed daily checkpoints

Enterprise

Every audit entry carries a cryptographic hash of its predecessor, so altering or deleting any entry breaks the chain; the head of the chain is signed daily and the checkpoint is emailed to the owner.

Administrators see the time of the last successful verification, and any integrity failure is reported.

How it works in the help centre

Audit export with independent verification

Enterprise

Exports are delivered as an archive containing the entries as CSV, the hash chain, the signed checkpoints, the public keys and a verification script.

The script (`node verify.mjs`) runs offline with no network access and reports either success or the first entry that fails verification. It can also validate an export against the fingerprint in a checkpoint email. Owner exports include entry content in the chain, so every line is verified.

Limits: Export files are available for 7 days, to the administrator who requested them.

How it works in the help centre

Where is data processed, and what reaches the AI provider?

Processing is confined to the regions the workspace selects, PII is detected and redacted per category before it leaves the workspace, and customer content is never used for model training.

Data residency (regional processing)

Enterprise

The workspace owner restricts AI processing to selected regions: the United States, China or Indonesia, with Europe to follow.

Requests are routed only to AI providers that process them in a selected region. There is no cross-region fallback: if no provider in the region can respond, the request is not sent. Web search, voice transcription and image generation, which run outside the region or cannot be confirmed to run within it, are disabled unless the owner enables each one. AI features on pages created in a restricted workspace remain bound to its regions wherever the page is opened. Administrators can request changes; only the owner can approve them.

Limits: Model availability varies by region; when no model is available, the request is declined rather than rerouted.

How it works in the help centre

PII detection and redaction

Enterprise

Messages and connector results are scanned for personal data before they reach the AI provider, and each category is set to allow, flag to the sender, or redact.

Categories include credentials and API keys, payment card, bank and IBAN numbers, government ID and passport numbers, names, email addresses, phone numbers, postal addresses, IP addresses and personal dates. Redacted values are replaced with consistent placeholders for the whole conversation and restored in the response, so users and tools still work with the real values. Detected values are never stored; only their position and category are recorded. If a scan cannot complete in time, the message proceeds with pattern-based redaction of credentials and the sender is notified. Voice-message transcripts are scanned like typed text.

Limits: Detection is not exhaustive, and company names are not detected. File attachments and documents read by the AI are not scanned. Content sent while redaction was disabled is not scanned retroactively.

How it works in the help centre

No training on customer content

Every plan

Conversations, files, memories, designs, voice recordings and data from connected applications are never used to train AI models or for advertising.

How it works in the help centre

Bring your own AI provider key

Depends on the plan

A workspace can route AI usage through its own AI provider account, under the contractual and data-processing terms it already holds.

Only owners and administrators can view or change the key. It is encrypted at rest, displayed only by its last four characters, and every change is recorded in the audit log.

How it works in the help centre

Data retention

Enterprise

Audit entries are retained without a cut-off, audit exports are removed 7 days after they are generated, and detected PII values are never persisted.

How it works in the help centre

How are identity, access and provisioning controlled?

Authentication is federated to the customer's identity provider on verified domains, multi-factor authentication can be mandated, provisioning is automated over SCIM, and access follows role-based controls.

Single sign-on (SAML 2.0 and OIDC)

Every shared workspace

Users on verified domains authenticate through the customer's identity provider using SAML 2.0 or OpenID Connect.

Administrators choose which verified domains use single sign-on and validate the configuration with a test sign-in. A domain can be set to SSO-only, which disables every other sign-in method for its accounts and terminates their existing sessions; SSO-only cannot be enabled before a successful test sign-in. Every configuration change, test and rejected sign-in is recorded in the audit log.

Limits: A domain can be bound to one Bluedeer workspace, and public email domains are not eligible.

How it works in the help centre

Domain verification

Every shared workspace

Domain ownership is established only by a DNS TXT record; an email address on the domain is never sufficient.

Records are re-validated continuously. If a record is removed, owners and administrators are alerted, and single sign-on for the domain is withdrawn unless the record is restored.

How it works in the help centre

Mandatory multi-factor authentication

Enterprise

Administrators can require multi-factor authentication for every member, the owner included: an authenticator app, emailed codes, a passkey, or single sign-on.

Non-compliant members receive a 7-day grace period, or none, after which their access to the workspace is suspended until they comply. No data is deleted.

How it works in the help centre

SCIM 2.0 provisioning and deprovisioning

Enterprise

The identity provider creates, updates and deactivates members automatically, and group membership determines each member's role and seat tier.

Limits: Group mappings can assign the administrator or member role, never ownership.

How it works in the help centre

Domain capture

Enterprise

New accounts registered on a verified domain join the organisation's workspace and request a seat, rather than operating as unmanaged personal accounts.

How it works in the help centre

Role-based access control

Every shared workspace

Every shared workspace has owner, administrator and member roles, and administrators decide which built-in applications and connectors the workspace permits.

Connector credentials are scoped to the individual user: no other member, administrators included, can view or use them. Members without a seat cannot use AI, run tasks or use connectors in the workspace.

How it works in the help centre

How is the platform itself secured?

Vendor support access is consent-based and audited, generated pages run in sandboxed isolation, inbound email is authenticated, and credentials are encrypted at rest.

Consent-based support access

Every plan

Each user controls whether Bluedeer support may access their account, which conversations support may read, and whether support may change settings.

Conversation access and settings changes are off by default. Support can be limited to read-only access, sessions last at most 30 minutes, and every support action is recorded, including in the workspace audit log.

Limits: Billing, plan changes, purchases, passwords, multi-factor settings, passkeys, sign-in methods and account deletion are never accessible to support.

How it works in the help centre

Sandboxed execution of generated pages

Every plan

Every version of a generated page is served from its own isolated origin, never carries the application session, and runs inside a sandbox.

Live pages that process the viewer's data cannot open pop-ups or new windows, and outbound links require the viewer to confirm the full address. Bluedeer's own analytics reject requests from page origins.

Authenticated inbound email

Every plan

Email forwarded to Bluedeer is accepted only from the account's own address or an approved sender, and only when the sending domain passes DMARC.

Automated senders are rejected.

How it works in the help centre

Encryption of credentials at rest

Every plan

OAuth tokens for connected accounts and AI provider keys are encrypted at rest, and keys are never written to logs.

Account protection

Every plan

Every account supports passkeys, multi-factor authentication by authenticator app or emailed code, and review and revocation of active sessions.

Repeated failed sign-in attempts are throttled to prevent credential guessing.

How it works in the help centre

Service continuity for Enterprise

Enterprise

An overdue Enterprise invoice never causes data loss: everyone in the workspace is notified 10 and 15 days after the due date, the owner and admins are warned by email on days 27, 28 and 29, the workspace becomes read-only on day 30, and payment restores full access immediately.

How it works in the help centre

For larger organizations

Contact sales

Bluedeer Enterprise adds identity governance with SCIM provisioning and multi-factor sign-in that can be required for every member, an audit trail, data residency and sensitive data masking in messages. From 50 seats.

Contact sales

See it in your own workspace.

Create a free account, then talk to sales about Enterprise for your organisation.