Products

Atarus Platform

One local program that will run our open-source security tools, driven from your own AI client or the command line, and keep the results as evidence.

In design Being designed. No code for it has been released.

The problem

Our scanners each have their own install, their own flags and their own output. Running several of them against one target and turning the results into something an auditor accepts is an afternoon of copying and pasting.

AI clients can now drive tools directly through the Model Context Protocol (MCP). Shell injection and tool poisoning are among the most common bugs found in MCP servers so far, so a server that runs security tools has to be built around those two from the start.

What it will do

Run atarus-recon, atarus-webscan, atarus-cloud, atarus-phishcheck and atarus-phpcheck from one place: as a local MCP server that your own AI client drives, such as Claude Desktop or Claude Code, or from the command line.
Hold every target to a scope you set. A target outside it will be refused, not warned about.
Keep each result with a record of what was checked and what was not, and turn results into a report through atarus-report-kit.
Take in Exposure Check and Extension Watch as modules once each of them is released.
Leave the reasoning to your own AI client and account, so your prompts and results never pass through us.

What it will never do

Pass text from the model or from a scanned page to a shell. Every tool will take typed arguments and check them.
Probe a system you did not place in scope, or start an active check without a confirmation the model cannot give for you.
Treat scanned content as instructions. Page text, headers and tool output will be wrapped and labelled as data before the model sees them.
Change a tool description silently. Descriptions will be fixed in the signed build, with their hashes published on this page.
Connect to anything other than the targets you name. Every connection it can make will be listed in the network table here.
Tell you a target is safe.

Reviews

Every product passes three reviews: an idea review before design, a design review before code, and a release review before it is listed here. A product that fails one is stopped, and the reason is recorded.

Idea
Not yet held

The plan that makes the platform the firm's main product was approved on 23 Oct 2026. No separate idea review has been held.
Design
Not yet held

The local server's design review, with its threat model, comes before any of its code.
Release
Not yet held

A test run of the built release and a security review of it, before it is listed here.

What this page will carry at release

The version number and its release notes.
A summary of the release review: what was tested, what was found and what was fixed.
A demo recorded against a lab target, never a customer system.
Every permission it needs.
A network table generated from the code, listing every connection it can make.
The hashes of the tool descriptions your AI client sees.
The hash that ties the source tag to the binary, published in two places.
The fingerprints of the signing keys.
The commands to check all of the above yourself.

Status and data

Status: In design. Being designed. No code for it has been released.
Released: Not yet. There is no download, version or price.
What it does: One local program that will run our open-source security tools, driven from your own AI client or the command line, and keep the results as evidence.
Data it collects: None. Scans, results and reports will stay on your machine.; A version for teams is planned for later and is not designed yet. Its first job would be a fleet view of Extension Watch. If a team admin deploys it, it would collect only extension IDs, versions, hashes, publisher, permissions and a device label. Never URLs, page content, cookies or identity.
Network: Connections only to the targets you name. Nothing else will be sent, and nothing will go to Atarus.
Source: The tools it runs are open source on GitHub today. The local version will be free, and how its own code is published will be stated here before its first release.

The rules it is built under

The same rules apply to every product we build. The security page covers them in full. The products page lists the others.

Auditor by default: Every tool observes, inventories and reports. Read-only is the default and needs no confirmation.
Active only on targets you placed in scope: Anything that probes a live system runs only against a target inside a scope you set, after a confirmation that software, including an AI model, cannot give on your behalf.
Never a false clean: Results say what was checked and what was not. No tool will tell you something is safe.
Threat model before code: Each module gets a written threat model at its design review, before any code.
Fuzzing on every parser: Anything that reads untrusted input gets a fuzz harness and limits on size, depth and paths.
Signed and reproducible builds: Releases will be signed, with separate keys for separate jobs, and built so that anyone can match the published source to the published binary.
Outside review once there is revenue: A security firm checking its own work is not an independent check, so an outside firm will review the platform once it earns.

Follow: release notes will go out in the research feed as posts of kind Tools, and our open-source code is on GitHub.