A standalone privacy boundary for AI work, and the egress control inside Haivian Factory.
The moment AI becomes genuinely useful on your own work, it needs your own work. Source, contracts, client material, unreleased research. What leaves your machine at that point is usually decided by whatever the tool defaults to.
For anyone holding confidential material or their own intellectual property, that default is the whole risk, and it is invisible.
Stages, inputs, outputs and decision points. Not the mechanism behind them.
Your original text, files and context remain on the device. Nothing is uploaded in order to be examined.
Protected material is identified locally, before anything is addressed to a provider.
Your policy sets the handling: allow, warn, seal, tokenize, redact, or block the request outright.
What reaches the approved provider is the protected form. The unprotected original does not travel.
The response returns without the provider having had access to the mappings held on your device.
Permitted values are restored on your machine, and a content-free receipt records the event, the policy and the authority.
If a remote model genuinely requires the hidden substance to perform the task, no privacy layer can both conceal that substance and let the same remote model use it. The safe choices are local processing, an approved surrogate, explicit disclosure or blocking the request.
A record of what left, when, under which policy and on whose authority, and the confidence that the rest did not leave at all.
Implemented and tested internally. Not generally commissioned.
Shield has no published evidence exhibits yet. When an exhibit is approved, its bounded question, method, artifact, result, verification route and limitations will appear in the journal.
We publish the evidence we have, not the evidence we expect. Nothing on this page cites a customer, a metric or an outcome that does not exist. Where a claim is architectural rather than demonstrated, the meter above says so.
Anyone whose working material is confidential or is itself the asset: professional services, research, legal and engineering teams who cannot simply paste the real thing into a provider.
Working on client material that cannot be disclosed to a third party
Protecting your own unreleased intellectual property while still using AI on it
Producing evidence of what was disclosed, for an obligation you have to meet
Local, on your own device
Organizational policy administration
Inside a Factory deployment
In the current local deployment, protected content and vault mappings remain on the device. The evidence ledger records the event, policy and authority, not the protected content itself.
Each model implies a different data boundary. See all four in detail.
Named rather than hidden. You can see the boundary, the behaviour and the evidence without being able to rebuild the mechanism.
All prices are in USD.
Pricing is published. Online purchase is not open; scope and availability are confirmed in conversation. Licence and engagement benefits are distinct from metered hosted compute and third-party provider spend.
The browser extension. Protects what you paste, on Chrome and Brave.
Confidential work, analysed by a cloud model that never sees the confidential part.
One policy across a whole organization, understood and enforced by Shield.
Shield inside Haivian Factory, so the disclosure boundary covers every assignment the Factory runs.
Licence and engagement benefits are distinct from metered hosted compute and third-party provider spend.
On its own, Shield governs what leaves your machine. Inside Factory it governs what leaves the whole system, so every assignment inherits the same disclosure boundary and every disclosure lands in the same evidence trail as the work that caused it.
The scoped Factory integration shown above is the connection for the stated program. It is not an unlimited-project licence.
Pricing is published. Online purchase is not open; scope and availability are confirmed in conversation.