Security — Causality.tools

Most security pages describe what a vendor promises to do with your data once they have it. This one is shorter than that, because the architecture removes most of the question: Causality Tools is a desktop application, and your case data stays on the machine that produced it.

Where your data lives

Causality Tools installs as a self-contained application. The database runs inside the application itself, the search component runs locally alongside it, and neither requires a container runtime, a server, or an account with us.

Adappt operates no multi-tenant service and holds no client case data. There is no tenancy of ours for your material to sit in.
There is no product telemetry. Nothing about a case, a question, a source, a graph label, a navigation path or an error leaves your machine — not opt-in, not anonymised.
Where several analysts need to work together, the shared component runs inside your own perimeter, under your access model and your operational controls. It is not a service we host on your behalf.

The AI models

The platform reasons with language models, and where those models run is a decision you make, not one we make for you.

Where the platform calls a hosted model, it does so with your credentials, under your own provider account and agreement. Model provider keys are held in your installation's settings store; they are masked in the interface and never reproduced in any exported record.
A deployment can be configured to use local models, so that no case content is sent to an external model provider at all.
Fetched page content is never shown to the model that decides what to do next. Research reads pages; the reasoning step sees only extracted, structured material. This is a security control as much as an accuracy one — hostile text on a web page cannot instruct the system.
Image payloads never enter a model prompt.

How the application protects itself

The desktop application is not merely a browser pointed at something. It ships with its own controls:

The macOS application is signed with an Apple Developer ID certificate and notarised by Apple.
The local API binds to loopback rather than to every network interface.
The local search component is fronted by a bearer token minted fresh on every launch.
The application will only use a data directory it owns, verified rather than assumed.
Archive contents are validated before extraction on restore, and link entries are refused.
Backups are written owner-only.

Isolation, where more than one case or team is involved

Case separation is structural rather than a permission check that could be forgotten.

Record identifiers are derived from the case they belong to, so the same real-world entity appearing in two cases is genuinely two different records. There is no shared row a missing filter could expose.
No case data can be opened without first resolving that case under the caller's own organisation.
The database enforces row-level security, with the application connected as a non-privileged role.
A request for a case belonging to another organisation returns the same response as a case that does not exist — it does not confirm that the case is there.

Where roles apply, they are re-read on every request rather than cached into a session, and sign-in is passwordless, through your own identity provider.

What the system can prove

The reason to trust an analysis is not that we say it is trustworthy. It is that the platform can show its working, on demand, as a document you can hand to a regulator, a court or an oversight board.

Decision lineage. For any conclusion, an exportable record of how the system got there: each step's plain-language reason, its confidence, the verbatim evidence behind it with its source, and which runs and which model produced it.
A transparency record per case. The case's declared purpose, the data actually used, what actually ran, and the safeguards in force — where every safeguard names the mechanism that enforces it. A safeguard that cannot name its enforcement is not listed.
These records are generated mechanically, not written by a model. A document whose purpose is proving honesty must not itself be able to invent. Where a record is incomplete, it renders as an honest gap rather than being filled in.

Human oversight

No output is acted on automatically where the consequence is material. Analysis, pattern detection and scoring are accelerated; the judgement, the approval and the accountability stay with a person, and the record shows who signed it off.

Certification and assurance

Adappt's delivery environment is certified to ISO 27001 and Cyber Essentials. Certificate scope is confirmed on request. Cyber Essentials Plus is in progress — we will say so until it is not.

What we ask of you, and what we have not built yet

We would rather state these plainly than imply protections we do not have:

Data at rest on the desktop relies on the operating system's full-disk encryption. We require it — FileVault on macOS — because the application does not yet encrypt its own store. Per-record encryption is planned.
Causality Tools ships today as a macOS application. Windows packaging is in development and is not yet code-signed; we will not ship an unsigned installer quietly.
Machine-level access is the security boundary on a single-user installation. Anyone with your unlocked machine has your cases.

Responsible disclosure

If you believe you have found a security issue, please contact us at security@causality.tools. We are grateful for responsible disclosure and will respond promptly. We do not currently run a paid bug-bounty programme, and we will not threaten anyone who reports a genuine issue in good faith.

For how deployment is scoped around your operating model, see Trust & Deployment. For how we handle personal data, see our Privacy Notice.