Security
Hive reads your numbers, your inbox and your documents, and acts across your tools. Every decision — the data model, the agent architecture, the approval flow — assumes the business is sensitive and consequential.
Safe by design, not by policy
Handing operations to software only works if the software is trustworthy by construction. Understand first, act second, prove it happened. Three things make that true: control over what Hive does, security over what it reads, and discipline over what it spends. The defaults are the conservative ones — you loosen them, never the other way around.
Acts when sure, asks when not
Work is scored by confidence and routed accordingly. Hive acts on what it is certain of, raises for approval what it is not, and escalates what exceeds policy. There is an approval on every action and a receipt for everything — what was done, the figures, the source trail. A master kill switch stops everything at once. Autonomy is earned per workflow; it is not assumed.
A firewall between reading and doing
Everything Hive reads — an email, a document, a web page — is treated as untrusted. An instruction hidden inside content cannot make Hive leak data or take an action it wasn't authorised to take, because there is a boundary between what it reads and what it can do. Prompt injection is ignored, export to an unknown destination is blocked, and a contact allow-list bounds who Hive can reach. Egress is locked.
Bends rather than breaks
A fallback at every step keeps Hive honest under real conditions. It knows two ways — by meaning, then by direct lookup. Live lookups are time-boxed and fall back to memory rather than hanging. It abstains instead of inventing, and self-corrects or admits it has no data. Risky actions route to a person, a slow model fails over to a reliable one, and when something does fail Hive surfaces the real reason rather than a vague error.
Your data, on your terms
- Default accessRead-only · scoped per tool
- ActionsNothing sent without approval
- CredentialsVaulted · rotated
- Workspace memoryYours · exportable · deletable
- Shared modelsNever trained on your data
- Every actionLogged · auditable · reversible
Spend tracks difficulty
Hive routes each task to the smallest model that can do it well. Fast models handle classification and extraction; mid-tier models handle low-stakes drafting; frontier models are reserved for hard reasoning. The same discipline that controls cost also limits how much capability is pointed at any one task, and a reliable model steps in whenever a faster one falls short.
Held to a high standard
Security and compliance are an active priority, not an afterthought. Trust is the precondition for handing over operations, so we hold ourselves to the standards below and are working toward formal certification now.
- SOC 2 Type IIIn progress
- ISO 27001In progress
- UK GDPRAligned
- Data Processing AgreementAvailable on request
- Encryption in transitTLS 1.3
- Encryption at restAES-256
- EU & UK data residencyIn progress
- Independent penetration testingPlanned
- Access controlLeast-privilege · audited
- Sub-processor listOn request
Security questions
For the legal detail, see the Privacy Policy and Terms. For anything else, or to request our sub-processor list, email hello@get-hive.ai.