Security × AI

Assess and build systems
from the attacker's perspective.

A founder who deeply understands both security and AI
handles assessment, design, implementation, and adoption end to end.

Get in touch View services
Who it's for
01

You are holding a core system with no specification and no one left who built it — too risky to touch.

02

The system can't go down, so assessment and remediation keep getting pushed back.

03

Physical security and cyber monitoring run separately, and no one is watching the seam between them.

04

A customer has asked you to show your security measures, but the legacy core system cannot be touched.

Why it matters

In an era where AI finds unknown vulnerabilities,
lacking security insight becomes a critical business risk.

Attackers use AI too. Vulnerabilities that were hard to find by hand are now exploited automatically and at low cost.

If the people building AI systems have no security perspective, those systems can't be defended. An engineer who understands both needs to be involved from the design stage.

Security
Penetration testing for financial institutions (GIAC GWAPT)
Started as a white-hat hacker at Nomura Research Institute. Consistently handled assessments meeting strict financial-sector standards such as PCI DSS and FSA guidelines.
Development
10+ years from requirements to production
Implementation experience spanning security, AI, and data platforms. The founder handles every phase directly — no external subcontracting.
AI
From machine learning through generative AI implementation
Supervised learning, causal inference, and reinforcement learning through RAG, AI agents, and MCP design. Evolved into generative AI while carrying the full ML context.
What We Provide

Investigate, fix, watch, respond.

Systems and facilities you cannot afford to stop: we investigate them, fix them and watch over them, from both the physical and the cyber side. You do not have to go in order. Order only what you need, one service at a time. The person who finds the problem fixes it.

ServiceWhat you receiveTermPrice (tax excl., guideline)
Assessment (investigate)Find what could stop the system, from the attacker's side. Report plus a one-page executive summary with estimated loss and a quote for the next stage2–4 weeksFrom JPY 500K
Specification analysis (investigate)Recover what a system with no documentation and no original developers actually does. Full specification and the scope of remediationquotedFrom JPY 3M
Remediation (fix)Fix it while it keeps running. The repaired system and a re-assessmentquotedFrom JPY 4M
Rebuild (fix)Rebuild it to stop less. The new system and an annual re-assessmentquotedFrom JPY 5M

Start with any service, or order just one. Integrated Monitoring (watch) is quoted after a site check. Already hit? See Emergency investigation → Full picture: Omamori →

Track record

Retail · Specification analysis, assessment and remediation

Nobody knows what is inside our core system.

Situation
A core system handling store sales, inventory and ordering. No documentation, and the contract with the original vendor had ended. If it stops, store ordering stops.
What we did
Analyzed source, configuration, logs and traffic to produce a full specification covering screens, batch jobs and external integrations. Then assessed from the attacker's side and presented risk and fix priority translated into business impact (ordering outage, tampered sales data). With the impact scope documented, applied fixes outside store hours.
Result
Delivered the specification and the scope of remediation, so the team now knows what breaks when they touch what. Critical and high findings fixed within 3 weeks and confirmed by retest.
Wholesale · Remediation

Orders, inventory and shipping all run on one internal system. We want it fixed, but if it stops, nothing ships.

Situation
A sales and inventory management system handling everything from orders to shipping instructions. A daytime outage means missing the day's dispatch.
What we did
For the findings from assessment (an unauthenticated admin screen, a single permission level for all users, API keys embedded for external integrations), set the order of fixes and remediated authentication, permission separation and audit logging without stopping shipments.
Result
Zero shipping downtime during remediation. Fixes confirmed by retest.
Manufacturing · Incident response

Someone got into our order and shipping system. We cannot issue shipping instructions. What happens to today's shipments?

Situation
An unauthorized login to the order and shipping management system. Entry point unknown, and unclear whether other servers were reached.
What we did
On the case the same day. Cut off the entry path and preserved logs and devices as evidence. Correlated authentication, network and cloud audit logs to identify the entry point and determine whether there had been any lateral movement. Set up interim procedures and an isolated environment so shipments could continue during the investigation. Rotated credentials, redesigned permissions and closed the way back in.
Result
Operations restored after 6 hours of downtime. No lateral movement or re-intrusion. Delivered an executive report and prevention measures.
Manufacturing · Assessment

If the production management system stops, the line stops. Is watching the cyber side enough?

Situation
A production-management core system and the server room inside the plant.
What we did
Mapped what could stop it, from cyber attack and physical intrusion to equipment failure and insider misuse, together with dependent systems and business impact, and assessed from the attacker's side. Reviewed server-room locks, entry records, and power and cooling monitoring through web-meeting interviews and documents, and identified the gaps between those and the cyber controls. Produced a one-page executive summary with estimated loss.
Result
One page that lets management decide what could stop the line and what to fix first. Critical and high findings confirmed fixed by retest.
Finance · AI agent assessment

We deployed an AI assistant over internal documents. The web app passed assessment, but is the AI side safe?

Situation
An internal document search assistant already in use. The web application had been assessed, but the AI-specific paths had not.
What we did
Inventoried the agent's tools, permissions and data sources, then ran attack simulations: prompt injection, retrieval of documents beyond user permissions through RAG, and privilege escalation through tool execution.
Result
Identified and fixed paths to documents beyond user permissions that the web assessment could not see. Retest confirmed that documents nobody should see no longer come out.
SaaS · Secure application development

We are building an internal AI agent on business data. We want it defensible from day one.

Situation
A new internal AI agent app connected to business data and tools.
What we did
Designed on zero trust: per-request authentication, WAF, network isolation, least privilege, audit logs. AI tool execution allowlisted, critical operations gated by human approval. Accompanied from design review through release.
Result
Every critical finding fixed before release. No security-related rework or downtime since release, and zero critical or high findings in external assessment.
Sound familiar? Get in touch Start with an assessment (from JPY 500K). The founder replies directly within one business day.
About the founder

Reo Onozawa

Former Nomura Research Institute / Founder, Green Tea LLC

Started as a white-hat hacker at Nomura Research Institute. After hands-on machine learning work, now supports the design and development of secure AI applications and the safe adoption of AI in the field.

Because the founder understands both security and AI, systems can be designed and implemented from an attacker's perspective. Every phase from requirements to production is handled directly.

GIAC GWAPT Registered Scrum Master
Full profile →
01
Security Engineer

Started as a white-hat hacker at Nomura Research Institute. Handled penetration testing and vulnerability assessment for financial institutions, consistently delivering assessment and remediation support based on OWASP, PCI DSS, and FSA guidelines. GIAC GWAPT certified.

02
Machine Learning Engineer

Applied supervised learning (demand forecasting, credit scoring, attrition prediction), unsupervised learning, time-series analysis, causal inference, and reinforcement learning in practice. Evolved into generative AI (RAG, AI agents, MCP) while keeping the full ML context.

03
Forward Deployed Engineer

Embeds into the client's organization, managing projects while handling development end to end. Supports assessment, design, implementation, and adoption.

Why choose Green Tea

The founder handles
every phase directly.

With a development vendor, members of varying skill levels join the project, so service quality is inconsistent. When the proposal team differs from the implementation team, what was agreed up front doesn't always make it into the build.

The founder handles every phase directly. No external subcontracting. From proposal to implementation, the person in charge never changes.

Development firm
  • Security, AI, and development split across a team
  • Handoffs occur between teams
  • Proposal owner differs from implementation owner
  • Staff change after the deal — or work is subcontracted out
Green Tea
  • Security × AI × development in one person
  • The founder handles every phase directly
  • Context never breaks from requirements to production
Integrated Monitoring "Omamori"

Watch both physical and cyber anomalies
on one screen.

Omamori uses your existing cameras and systems as-is, ties physical intrusion and cyber anomalies onto a single timeline, and hands your team the recommended response. People decide what gets executed.

From detecting a combined attack to the recommended response (about 1 min)
Learn about Omamori →

Reach out anytime.

It's fine if you're not sure where to start. The founder responds directly by the next business day.

Get in touch