The core principle: the AI proposes, rules decide
Hive’s AI never holds your credentials and never executes anything directly. Every action follows the same path:- Proposals are typed. The model can only propose actions from an allow-list of known action types, with structured parameters — not free-form commands.
- Policy is code, not a model. A deterministic engine checks every action before it runs: the kill switch, spend caps and rate limits always apply, and any work that uses credits is checked against the credits it would draw on before it runs (on Business, the paying member’s own; on an earlier plan, the workspace’s pool), failing closed if the check cannot be made. In Ask and Signals it also weighs the action’s autonomy rung, its blast-radius tier and confidence; everywhere it weighs the account’s Read, Draft or Act level, an agent’s or a workflow’s selected actions included. An action selected for an agent or a workflow runs without an approval card only on an Act account. It waits for your approval when where it sends did not come from words you trust, and the few actions that always wait for a member’s click, such as a LinkedIn post, still become a card. A workspace limit can stop it, when Hive refuses it and says why. See What agents and workflows do by themselves.
- Credentials stay with the executor. The component that talks to your tools is separate from the model and is the only part that can use a credential.
Isolation between workspaces
Each workspace’s data is isolated from every other workspace at more than one layer:- Every record carries its workspace id, taken from your verified sign-in session — never from a request body or from model output.
- Postgres row-level security is enabled and forced on workspace tables, so the database itself refuses to return another workspace’s rows.
- Application code also filters by workspace on every query, as a second, independent check.
- Requests for another workspace’s records return “not found”, so ids cannot be probed.
Identity and access
- Sign-in is provided by Clerk, a dedicated identity provider. The sign-in options available to your workspace are those configured for Hive’s sign-in.
- Roles are held by Hive. Your identity provider says who you are; Hive’s own member record decides what you can do. See Roles and permissions.
- Sessions are visible to you. Settings → Account → Security lists your active sessions, with sign-out per device.
- AI clients sign in as you. A client connected to the Hive MCP server uses OAuth 2.1 with PKCE, ten-minute access tokens and rotating refresh tokens, gets only the per-area read or write access you approve (read-only unless the client asks for more), and has your role checked on every call. It cannot add an address, number, link, channel or repository to an agent or a workflow, or approve an action Hive held because of where it sends: those stay in the Hive app. Settings → Connected AI clients lists and revokes them.
- Private by default. Chats, workflows, agents and files belong to their creator until shared, and sharing one never shares the owner’s connected accounts.
Protecting connected accounts
- Credentials are encrypted per workspace. Connector and model secrets are sealed with AES-256-GCM using a key derived for each workspace, bound to that workspace’s id, so one workspace’s secret cannot be decrypted in another’s context.
- Secrets are never shown to the model and never logged.
- OAuth tokens for managed connectors are held by Hive’s connector platform, not in Hive’s application tables.
- Each account’s owner sets what Hive may do with it. A new account starts on Act: Hive takes the action and tells you what it did, with undo where it can. An action that cannot be undone, or one a safety check holds back, still comes to you as an approval card. Its owner can move it to Draft (Hive prepares each action for your approval) or Read (search and summarise only), and an Owner or Admin can cap a provider for the whole workspace. These levels hold in Ask and Signals and for an agent’s or a workflow’s selected actions alike. See Accounts and autonomy.
Defending against manipulated content
Hive reads emails, documents, tickets and web pages written by people outside your business. Any of them could contain text trying to steer an AI. Hive assumes they might:- Untrusted content is marked as untrusted in every prompt (“spotlighted”), and treated as data, never as instructions.
- Knowledge keeps its trust level. Facts carry a trust tier, and anything derived from a low-trust source stays low-trust.
- Know who Hive may send to. An email can go to any address your connected account can send to. In Ask and Signals an email cannot be unsent, so on Draft or Act it becomes a High risk approval card that names each recipient and where it came from; check them before you approve. The one exception is a reply you owe in your own Gmail or Outlook thread, which Signals may send by itself on Act only when Hive read nothing beyond that thread to write it (see Email replies on an Act account). An agent’s or a workflow’s selected send goes out with no card only from an account on Act (on Draft it waits for approval), when its recipient came from words you trust, such as the agent’s instructions or a value you typed to start the run. A recipient that came from what the run read, such as an address in an incoming email, waits for your approval instead: see Where an action sends. For a person to check every send of a workflow, put an Approval step before it. A Slack direct message still goes only to a workspace member. See Know who Hive may email.
- The model has no general-purpose fetch or execute tool. It can only use vetted, typed tools.
- What can reach the open web. Hive’s web search and page reading are not held back after an answer has read your private data: a search query may include what the answer read, and Hive’s search provider receives it. Hive opens only pages a search returned, links you wrote and same-site links on pages it opened — never an address the model made up — and answers never load images from links. A workflow’s automatic Web search leaves out what came from your connected apps. Other outside lookups, and MCP tools that open whatever address they are given, are not called after a private read. See What can leave Hive after a private read.
- External MCP tools are pinned. When you connect an MCP server, Hive records a fingerprint of each tool’s name, description and input schema. A tool that changes later, or drops the server’s read-only mark, is not silently trusted. Every MCP tool is treated as high-risk and not read-only unless an operator decides otherwise; Ask calls one on its own only when the server marks it read-only and its name is plainly a read, and everything else waits for your approval. A write tool an agent’s author selected runs in that agent’s runs without a card, as its other selected actions do. In runs with nobody there, such as scheduled agent runs and workflow steps, links in a tool’s result that open a file for anyone holding them are removed before the AI reads it. See Using MCP tools in Ask.
Approvals, receipts and undo
- Approvals are bound to the exact action. An approval covers the precise parameters that were reviewed and expires after 7 days. A changed action needs a new approval.
- Every action leaves a receipt in the Audit log: what was noticed, what was planned, how it was decided and what happened.
- Reversible actions can be undone from the audit log.
Stopping everything
The kill switch halts every autonomous action and approval in the workspace immediately, survives restarts and fails closed. Circuit breakers on action rate, spend and error rate trip automatically. See Safety and control.Signed webhooks
Every webhook Hive receives is signature-verified — from its payment processor, identity provider, connector platform and job queue, and from your own incoming webhooks, which use HMAC-SHA256 signatures over the raw request body.What we do not claim
We would rather be precise than impressive. Hive does not currently publish third-party certifications or attestation reports. If your organisation needs a security questionnaire completed, contact us.Responsible disclosure / contact
If you believe you have found a security vulnerability in Hive, please report it privately through Get support, with enough detail for us to reproduce it. Please do not access other customers’ data or disrupt the service while testing, and give us reasonable time to fix an issue before disclosing it publicly.Related
Data protection
Retention, deletion, sharing and analytics.
AI and models
Which AI providers Hive uses and how.
Infrastructure
Hosting, regions and providers.
How Hive works
The architecture end to end.