Concept app · v0.2.0 · Notarized

Lynceus

A security review of everything that runs through your Jamf Pro: scripts, extension attributes, policy commands and script parameters, local accounts in policies, configuration profiles and package scripts. Lynceus reads them live with read-only API requests or offline from Gimle backups. A local AI reviews each object first, then 39 deterministic rules confirm, add and check what the model claims. You get an executive summary, a security report and a fix list for engineers.

Concept app. Lynceus explores whether a small local AI, kept honest by deterministic rules, can give a Jamf Pro team a useful security review of everything it deploys. Version 0.2.0 is signed and notarized. It has been tested in a demo environment only, not against a production Jamf Pro, so treat its findings as a starting point for your own review. Named after the Argonaut who could see through walls.

Download Lynceus (2.7 GB) Read the Docs

Lynceus showing the findings of a second scan of the demo tenant Demo Corp: risk grade E, 18 high and 10 medium findings, 2 new and 6 fixed since the last scan, and a hardcoded credential with its masked code in the inspector

Features

See through every script you deploy

🔍

Everything That Runs

Scripts, script-type extension attributes, policy Files & Processes commands, script parameters $4–$11 and local accounts, computer and mobile device configuration profiles, and pre/postinstall scripts from packages in Gimle backups.

🧠

AI First, Rules Second

The bundled Gemma 4 model reads each object like a reviewer. Then 39 rules (30 security, 6 hygiene, 3 deprecation) run. A finding both agree on is confirmed. An AI finding that quotes code which doesn't exist is marked unverified and doesn't count. The AI alone never rates a finding critical.

👁️

Read-Only

Live scans only send GET requests, through Gimle's backup engine, so a read-only API role is enough. Or skip the API entirely and scan a Gimle backup, encrypted ones included.

🔒

Secrets Masked

Passwords, tokens and keys found in your content are masked everywhere Lynceus stores or shows them: findings, history, logs and every report. The tests check the reports for leaks.

📑

Three Reports

An executive summary (risk grade, top risks, trend), a security report (every finding with evidence, coverage, accepted risks, the rule set used) and a fix list for engineers, as PDFs, plus CSV and Markdown.

🔁

New and Fixed

Every scan is kept, so each run shows what's new and what's fixed. Unchanged objects aren't sent to the AI again. Accept a risk or mark a false positive with a reason; accepted risks reopen after 180 days by default.

📍

What's Actually Deployed

Per object: is it in use, which policies use it, and how many devices it reaches. First seen and last changed come from the Gimle backup history. Filter to what's in use and fix that first.

✅

Triage in Bulk

Group findings by object or by finding, select several with ⌘-click, ⇧-click or ⌘A, and accept, dismiss or reopen them in one step.

🎯

Scope Before a Scan

Skip an object or check it with rules only, each with a reason. The scope applies to every later scan, and the reports list what was left out. A skipped object never counts as fixed.

📐

Your Own Rules

Turn built-in rules off or change their severity, write custom rules as regular expressions with a live test, and share them as JSON. MDM can deploy one rule set and lock editing.

⏱️

Live Progress

Watch a scan: progress, time left, the object under review and its stage, and the findings so far. If the AI fails, the scan pauses and retries every minute, and a stopped scan resumes where it left off.

🛡️

Local AI, Managed Options

The AI runs on the Mac by default. If a larger model is worth it, an admin can point Lynceus at the company's own OpenAI-compatible server through MDM. Secrets are masked before anything is sent, and AI can be turned off for rules-only scans. No telemetry.

See It In Action

From scan to fix list

Lynceus scanning Demo Corp: 6 of 20 objects done, the model reading the current script, the next objects in the queue, and the findings found so far

Scan running: progress, time left, the object the model is reading, and the findings found so far.

Lynceus findings grouped by object, with in use, device counts and new markers

Grouped by object: findings per script, policy or profile, with whether it's in use and how many devices it reaches.

Lynceus scope sheet listing every object with Scan, Rules only or Skipped and a reason

Scope: scan, rules only or skip, per object and with a reason, for every later scan.

Lynceus Rules window with a custom rule matching an old update server next to the 39 built-in rules

Rules: the 39 built-in rules and a custom rule with its regular expression and object types.

Lynceus executive summary PDF with risk rating, summary, an analyst note written by the local AI and the top risks

Executive summary: risk grade, changes since the last scan, top risks in business terms.

Lynceus engineer fix list PDF showing a hardcoded credential with the masked code line and how to fix it

Fix list: every finding with the code in context (secrets masked), why it matters and how to fix it.

All screenshots show a made-up demo tenant, "Demo Corp". No real organization data.

Status

Where things stand

System Requirements

macOS 27 or later, Apple Silicon only.

The bundled model needs about 2.7 GB of free memory; on an 8 GB Mac, close other apps before a scan.

For live scans: a Jamf Pro API client with a read-only role.

Build Status

Version 0.2.0. Signed with a Developer ID certificate and notarized by Apple. The 2.7 GB package includes the Gemma 4 E2B model (Apache 2.0).

54 unit tests pass in 17 suites (Swift 6, strict concurrency).

What's Verified

Tested in a demo environment only, not against a production Jamf Pro. The AI review was checked against the real bundled model on demo scripts. Pausing and retrying when the AI fails was tested end to end with a local fake server.

Why It Exists

Years of scripts pile up in every Jamf Pro: passwords in clear text, downloads run as root, security controls switched off. Lynceus finds them and shows which ones still reach devices.