ASTRLINK GUARD

Privacy,
detected locally.

AstrLink Guard is AstrLink's own lightweight privacy model.
It finds names, phone numbers, keys and other sensitive data,
and redacts them entirely on your machine, with no network access.

Scenario demo
On your computer At the AI provider
1What you give the AI

Tidy up these deployment notes: server 10.0.3.17IP address, log in with api_key = "q7Vt2LmX9pRk…Secret", and ask Jane DoeName at +1 415-555-0132Phone if anything breaks.

AstrLink Guard found 4 things to protect, locally

2What the AI receives

Tidy up these deployment notes: server protected address, log in with api_key = "protected secret", and ask a colleague at protected phone if anything breaks.

The AI only sees placeholders and does the task as usual.

3The AI's reply

Server: protected address
Credentials: api_key = "protected secret"
Contact: a colleague, phone protected phone

4What you seePrivacy restore

Server: 10.0.3.17
Credentials: api_key = "q7Vt2LmX9pRk…"
Contact: Jane Doe, phone +1 415-555-0132

Placeholders become the originals again. You get the full result; the AI never saw these values.

Fictional content. Actual handling depends on the policies you enable and what is detected.

01 — Why you'd want it

Rules see patterns.
The model reads context.

AstrLink's built-in rules need no download and spot emails, phone numbers, card numbers and common key formats by their shape. But much private data has no fixed shape: a person's name, an address, an account, or a token buried in a log. AstrLink Guard is for those.

Works out of the box

Built-in rules

Available the moment you install, fast and always on.

  • Emails, phones and card numbers
  • Keys and tokens in common formats
  • No extra memory
Local privacy model

AstrLink Guard

Judges by context, so it recognises things with no fixed format.

  • Names, addresses and accounts
  • Credentials inside code, configs and logs
  • Chinese, English and mixed text

Pick one detector in Security policy: regex rules, or the local model. Switch to the model when you want finer protection.

02 — What it recognises

More than
just patterns.

Ten kinds of everyday sensitive data, in one pass. Order numbers, amounts, version numbers and hashes aren't private, and aren't treated as such.

03 — Why it's safe to use

Stays on your machine,
doesn't slow you down.

The model is verified and run on your computer. Request bodies are never sent to the model's source.

Local only

No network access at inference. Your content never leaves the machine to be checked.

Small and fast

The lightweight build is about 35 MB and checks a paragraph in tens of milliseconds, so your agent keeps moving.

Open source

Apache-2.0. The model, its evaluation and usage are all public on Hugging Face.

04 — How to turn it on

Three steps
to put it to work.

Installers don't include model files. Rules work right away; download the model when you want it.

  1. 01

    Open Security policy

    In AstrLink, open Security policy and switch Detector to the local model.

  2. 02

    Download AstrLink Guard

    In the model library's Built-in list, pick AstrLink Guard and a variant: FP32 for the best accuracy, INT8 for a smaller, faster build. Downloads pause and resume, or import model files you already have.

  3. 03

    Dry-run, then enable

    Run a dry run to see what gets detected, then choose the request action: warn and continue, block the request, or redact and continue. With response restore on, placeholders in the AI's reply and in tool-call arguments are swapped back locally.

If the model can't load, requests are rejected rather than sent unchecked.

05 — Where it stops

One layer of protection,
not a guarantee.

Every detector misses some things and flags others. Check the results with a dry run before relying on it day to day.

Parameters
37M
INT8 build
35 MiB
FP32 build
141 MiB
Languages
Chinese · English
License
Apache-2.0
Version
0.1.0
Full model card: architecture, evaluation and usage
Put it to work

Download AstrLink, keep privacy local.

Rules work out of the box. The privacy model downloads on demand.