村正
村正 Forged for autonomous offensive security

Every strike proven.
Every blade yours alone.

Like the legendary blades of Sengo Muramasa, this platform is built for one purpose: fully self-hosted AI agents that plan, execute and document penetration tests across your web apps, mobile builds and infrastructure — every finding backed by tamper-evident proof you can verify yourself.

Request early access See it in action
Self-hostedYour infrastructure, your data
Evidence-firstSigned proof for every finding
Web · Mobile · InfraFull-surface coverage
Bring your own modelsClaude, GPT & any compatible endpoint
muramasa — security operations overview
Muramasa security operations dashboard showing usage trends, active engines and recent assessments
Capabilities

Everything a modern security team needs,
in one hardened platform.

Purpose-built for continuous, verifiable security testing — from first scan to verified fix.

Autonomous attack agents

Independent agent teams plan and execute their own multi-round static and dynamic attacks against authorized targets — cross-checking each other so one model's blind spot never becomes a missed vulnerability.

Tamper-evident evidence vault

Every exploit path produces signed, append-only proof receipts with independent integrity verification — findings close only after explicit retest verification.

Mobile SAST & DAST

Upload Android or iOS builds and select static, dynamic or combined modes on private device runners — including replacement-build retests after fixes.

Finding review & severity validation

Independent reviewer models validate severity before publication, deduplicate candidates, and preserve original ratings alongside operator risk decisions.

CVE & advisory watch

Versioned system memory plus OSV, NVD and CISA feeds produce explainable advisory matches for registered software — with bounded automatic verification.

Security copilot & AI red teaming

A built-in assistant answers questions across workspaces while dedicated red-teaming workflows probe your own AI deployments for prompt injection and more.

Developer fix loop

Developer fix submissions automatically queue focused retests, retain change history, and open verified pull requests through optional GitHub integration.

Post-quantum readiness

Dedicated monitoring tracks cryptographic posture across your estate so you can plan the migration to PQC before it becomes an audit finding.

Usage analytics & cost control

Per-million-token input/output pricing, USD/JPY switching and live usage charts keep model spend visible and predictable across teams.

Architecture

The entire system on one map.

Follow one assessment end-to-end — from who gets in, through sealed execution cells, to a fix that only closes when the original exploit fails on retest.

Muramasa // System Architecture

self-hosted control plane · disposable execution cells · in-house models

01 Who gets in

Corporate SSOsingle sign-on · step-up verification
Least privilege rolesoperator · developer · viewer
Hardened edgeencryption-only · CSRF defenses

02 Central control & audit log

Operations consoletargets · campaigns · findings
Immutable ledgerevery action, replayable forever
Evidence vaultcourt-grade encrypted proof
Threat radarglobal advisories vs inventory
Verdict engineexploitable risk ≠ scanner noise

03 Sealed execution cells

Priority queuerisk-ranked · lease-based tracking
Disposable containersshredded on completion
Vanishing credentialsminted per job · revoked at end
Guarded rolloutscanary releases · instant rollback

04 The six offensive engines

Web engineinjection · auth flaws · misconfig
Attack plannerchains flaws into full paths
Operator agent αautonomous offensive toolkit
Operator agent βindependent cross-check
Mobile labstatic + runtime teardown
LLM red teamjailbreak · injection · leakage

05 In-house AI models

Private GPU clusterweights pre-staged · fast restart
Isolated namespacefixed quota · short-lived tokens
10-point validationno traffic until all checks pass
Zero external callsnothing leaves your perimeter

06 Finding → fix → re-test

Registertarget + context
➜
Attacksix engines run
➜
Confirmseverity validated
➜
File issuewith repro steps
➜
Patchdeveloper ships fix
➜
Re-testoriginal engine verifies
➜
Closeonly if exploit fails
村正 MURAMASA SIGNAL ACTIVE
AI infrastructure web mobile
0independent offensive engines sweeping your estate
0external AI calls — inference stays in-house
0%of closures require a passing retest
0+backend regression tests passing
How it works

From target registration to verified fix — one loop.

The normal workflow is a closed cycle: every remediation is proven by a focused retest before anything is marked resolved.

muramasa — system recon graph
Interactive system reconnaissance graph mapping discovered hosts, ports and services

Continuous visibility, retained context.

Muramasa keeps system inventory and advisory memory between assessments, so each new scan starts from everything learned before — not from zero.

  • Register targets: URLs, hostnames, IPs, Git repos, source ZIPs, APK/IPA builds
  • Attach HLD/LLD, requirements and purpose documents as encrypted context
  • Queue assessments with priority, scheduling and global concurrency controls
  • Review findings with independent severity validation before any ticket ships
  • Close only when a focused retest proves the fix works
Inside the product

A scorecard, not a scare deck.

This is a real executive scorecard generated from live platform data — every numeral is deterministically aggregated from retained records, never written by hand or invented by a model. Full evidence inventory, artifact checksums and limitations ship with every report.

muramasa — reports & downloads
Security report preview with finding records and export options
Pricing

Simple plans for serious security.

Work in progress — pricing below is provisional and subject to change.

Team

For small security squads getting started with autonomous testing.

—/ monthTBA
  • Up to 3 concurrent assessments
  • Web & repository scanning
  • Evidence vault & report exports
  • Email support
Most popular

Enterprise

Full-platform deployment for organizations that test continuously.

Contact usTBA
  • Unlimited concurrent assessments
  • Mobile SAST/DAST + private device runners
  • AI red teaming & PQC monitoring
  • GitHub automation & verified fix PRs
  • Priority support & deployment assistance
Talk to us

Self-hosted / OSS

Run the entire stack on your own hardware with your own models.

$0licenseTBA
  • Complete Docker Compose deployment
  • Bring your own LLM endpoints
  • All engines included
  • Community support
FAQ

Questions, answered.

Where does my data live?
Your browseroperator console
→
Your serverportal · queue · database · evidence vault
→
Your GPU clustermodel inference
No third-party cloud appears anywhere in this chain.

Muramasa is fully self-hosted. Every component above runs on infrastructure you own, and no finding, target detail or credential ever leaves your perimeter.

Do I need GPUs?
8vCPU cores
32 GBRAM
250 GBSSD storage
0GPUs required

This single server runs two concurrent full assessments. Heavy model inference is delegated to your separate LLM endpoints, so the platform itself remains inexpensive to host.

Am I locked into one AI vendor?
Anthropic · Claude (Messages API) OpenAI · GPT (Chat Completions / Responses) Any self-hosted gateway

All three are first-class citizens — there is no vendor-model requirement anywhere in the platform. Credentials are encrypted at rest, each assessment receives only its own scoped keys, and providers can be swapped in minutes.

How are false positives prevented?
  1. Cross-examinationIndependent reviewer models re-validate every severity rating before publication.
  2. DeduplicationCandidate and duplicate classifications are preserved — one real issue never becomes fifty tickets.
  3. Proof gateNothing closes until a focused retest makes the original exploit fail on record.

Put your security on autopilot.
Keep the evidence manual.

We're onboarding early design partners now. Tell us about your stack and we'll get you running a real assessment in days, not quarters.

Request early access Explore features
村 正