Security

You decide what leaves your building.

Core runs on a machine you own. Below is every connection it makes, what it sends, and how to switch it off.

  1. 01

    One firm, one machine

    Each Core serves a single firm, on hardware that firm owns. There is no shared system and no other customer on it.

  2. 02

    Nothing phones home

    Core sends no telemetry, analytics or error reports, and never updates itself. Updates run only when your administrator runs them.

  3. 03

    Encrypted to your machine

    TLS ends on your machine. When you reach Core from outside the office through our relay, the relay passes encrypted traffic through and cannot read it.

  4. 04

    Your text can stay local

    With a local model and on-machine indexing, mark a folder on-premises only and none of its text leaves. If the local model is down, Core skips the step rather than fall back to the cloud.

  5. 05

    Agents ask before they act

    In chat, the agent reads your library and, if you connected them, searches your research portals. Its database queries are read-only. Portal downloads and new sources wait for a person to approve them.

  6. 06

    Answers you can audit

    Each answer is logged with the model that wrote it, where it ran and its review status, and the log exports as a spreadsheet.

Outbound

Every connection Core makes.

Nothing else leaves the machine. Most of these happen only after you switch them on.

  • Cloud model (Anthropic, OpenAI or Google)

    Your question and the passages used to answer it. Without on-machine indexing, also document text and search queries for the index.

    When
    If a cloud model key is set
    To turn it off
    Use a local model and set EMBED_ON_HOST=1

  • Your own model server

    Prompts, on your network

    When
    If you run one
    To turn it off
    Unset ONPREM_LLM_BASE_URL

  • Yahoo Finance

    Ticker symbols in your book and watch list. No positions or quantities.

    When
    Four times a day
    To turn it off
    Remove the tickers

  • Frankfurter (ECB rates)

    Currency codes and a date range

    When
    Four times a day
    To turn it off
    Remove the tickers

  • SEC EDGAR

    Company identifiers for US tickers

    When
    If you add SEC filings as a source
    To turn it off
    Remove the source

  • Websites you add

    A request for the page

    When
    If you add a site or page
    To turn it off
    Remove the source

  • Microsoft 365, Google Workspace

    Sign-in tokens, for a read-only sync

    When
    If you connect them
    To turn it off
    Disconnect

  • Research portals (J.P. Morgan, Visible Alpha)

    Company names, under your firm’s login

    When
    If your login is set
    To turn it off
    Remove the login

  • Web search, through Anthropic

    Search queries about a company

    When
    When someone runs source discovery
    To turn it off
    Leave the cloud key empty

  • Hugging Face

    A download request for the indexing model

    When
    First start of on-machine indexing
    To turn it off
    Pre-load the model

  • GitHub (model library)

    A request for a public list of model prices. Nothing about you.

    When
    When Core starts
    To turn it off
    Set LITELLM_LOCAL_MODEL_COST_MAP=True

  • Treasurer release channel

    A version check and the update

    When
    When your administrator runs an update
    To turn it off
    Do not run it

Data flow

What leaves, machine by machine.

Three ways to run Core. Pick one to see what stays in your office and what goes to a cloud model.

FIRM NETWORK · ON-PREMISESINTERNETSOURCEResearch filesPDF · DOCX · emailSOURCEPortfolio bookpositions · CSVSTORAGE · LOCAL DISKLibrary + indexSQLite · Qdrant · Neo4jCLIENT · LANAnalystsbrowser · HTTPSMODELNo local modelinference in the cloudCoreYOUR MACHINEIngestparse · OCR · dedupeIndexembeddings · knowledge graphAgentsretrieve · cite · check figuresEGRESSTLS 443THIRD PARTYCloud modelyour own agreementLeaves: Questions, passages, index textstays on-premleaves the firm

Tightest setup

Run Core with nothing leaving.

  1. 01HardwareAn NVIDIA DGX Spark or RTX PRO workstation, with an open model running on it.
  2. 02ModelPoint Core at that model and leave every cloud model key empty.
  3. 03IndexingSet EMBED_ON_HOST=1 so documents and searches are indexed on the machine.
  4. 04FoldersMark sensitive folders on-premises only.
  5. 05Model weightsPre-load them, so the first start needs no download.
  6. 06DiskTurn on full-disk encryption: FileVault, BitLocker or LUKS.

In detail

For your security review.

Storage
Files, the index and answers live on the machine’s own disk. The database, search index and cache listen only on the machine itself.
In transit
HTTPS to Core, with TLS ending on your machine. TLS to every outside service.
Credentials
Passwords and keys for connected services are encrypted at rest.
Accounts
In deployments with user sign-in, each person has their own login. Passwords are hashed with Argon2id, sessions expire, and two-step sign-in is required for admins.
Roles and data access
In those deployments, users and admins are separate roles. Each person sees only the funds they are cleared for, enforced in the database, and access fails closed.
Sign-in records
In those deployments, sign-ins, failed attempts, password and two-step changes, and account changes are logged.
Web content
Downloaded pages open in a sandbox. Requests to private or internal addresses are refused.
Deletion
Deleting a document removes the file and everything derived from it.
Your data and models
Each firm’s Core is its own. Your data never trains a model for anyone else.
For your regulator
One download: the model card, controls mapped to the Hong Kong SFC’s four principles for generative AI, the answer log and validations.

Report a vulnerability

Found a security issue in Core or treasurer.net? Email us with “Security” in the subject.

contact@treasurer.net

Bring your security team.

We will walk them through the machine, every connection on this page and the inspection pack.

Request a demo

© 2026 Treasurer Pte. Ltd. All rights reserved.

Member of the NVIDIA Inception Program