01 / CHALLENGE THE MODEL
AI red teaming & model assessment
Test adversarial inputs and multi-turn manipulation to assess prompt injection, unauthorized responses and sensitive-information leakage in AI applications.
CHENGDU, CN / AI SECURITY LABRESEARCH · ENGINEERING · OFFENSE
OFFENSIVE THINKING. STRONGER BOUNDARIES.
Before the attack,
find AI’s weak points.
As AI learns to reason, decide and act, security must move with it.
With AI security at our core, Yofune Labs applies intelligent offensive testing to models, agents and business systems to uncover hidden risks and validate their real-world impact.
ADVERSARIAL
INTELLIGENCE
INTELLIGENCE
01—03BOUNDARY ANALYSIS
Exploring unknown boundariesGENERATIVE FIELD
Every intelligent action deserves verification.SCROLL TO EXPLORE
01THE NEW ATTACK SURFACE / CAPABILITIES
Greater capability.
Deeper scrutiny.
INTELLIGENCE NEEDS A COUNTERFORCE.
From prompts to tool calls, from code to business permissions. We take assessments to where AI actually operates, testing boundaries through adversarial research and supporting decisions with evidence.
02 / CONTAIN THE AGENT
Agent & tool-chain security
Trace agents from task interpretation to tool execution. Identify privilege expansion, context contamination and unintended actions to establish controlled operating boundaries.
03 / VALIDATE THE SYSTEM
AI-powered security validation
Run controlled adversarial assessments of models, agents and business systems. Combine attack-path analysis with evidence review to validate security boundaries and guide remediation and retesting.
MODEL AGENT APPLICATION ENTERPRISE
As systems evolve, our validation goes deeper.
02THE ARSENAL / PRODUCTS
Offensive thinking.
Practical tools.
BUILT TO FIND WHAT OTHERS MISS.
Our focus on AI security and intelligent offensive testing turns research into products for real workflows, helping security teams discover risks, run controlled validation and improve continuously.
Product overview/01
AGENTIC OFFENSIVE SECURITYKUROSHIOAgentic security
Put agents on the security front line.
An AI agent platform for validating complex systems. Connect objective decomposition, attack-path reasoning, controlled testing and evidence collection—so every investigation produces findings your team can review.
- Multi-agent collaboration
- Attack-path reasoning
- Scope and policy controls
- Evidence-backed reporting
KUROSHIO / AGENT OPERATIONSDEMO SCENARIO
MISSION / 024
Test scope constrainedExternal attack-surface validation
TASK → AGENTS → EVIDENCE
Planning agent / Path reasoning02 / 04
Break down validation goals using asset context and produce reviewable attack-path hypotheses.
01Establish target context
02Break down candidate paths
03Submit validation plan
DeliverableAttack-path hypothesis
Agents find signals. Evidence informs judgment. Your team decides.HUMAN IN CONTROL.
03OFFENSE WITH PURPOSE / OUR APPROACH
Untested security
is only an assumption.
We look beyond whether an anomaly can be triggered: how it happens, what it affects and how to stop it. Security research becomes a starting point for continuous improvement.
About 17 seconds · 4 stages
Interactive illustration · No real attacksFROM ASSUMPTION
TO EVIDENCE.
TO EVIDENCE.
Complete validation
goes beyond finding problems.
ScopePathEvidenceImprove
VALIDATION LAB
Demo ready
STAGE 01 / 04
01First, know what you are protecting.
ENTERPRISE AI / ASSET MAPIN SCOPE
Explicitly authorized test environment
Foundation modelMODEL ENDPOINT
Business agentAI AGENT
Enterprise knowledge baseRAG CONTEXT
Tools & permissionsTOOL ACCESS
Boundaries before actionsConfirm assets, identity relationships and data flows together.
04ASSETSMap models, agents, knowledge bases and tool permissions together to define the test scope and trust boundaries.
SCOPE MAP4 asset types / 1 authorized scope
Assets and permissions mapped · Scope confirmed
Press play or select a stage above0.0s / 16.8s
04YOFUNE, ARIAKE / ABOUT US
YOFUNE
ARIAKE
ARIAKE
FROM THE UNKNOWN,
INTO CLARITY.
Chengdu Yofune Ariake Technology Co., Ltd.Into the unknown.
Into the unknown.
Toward clarity.
Yofune—a night vessel—is how we approach the unknown.
Ariake—the coming light—is what we expect technology to bring.
We are a technology company with AI security at our core. For us, Yofune Ariake means exploring while exercising sound judgment: entering complex systems, finding overlooked signals and turning uncertainty into security questions that can be understood and tested.
As AI moves from answering questions to taking action, we work at its boundaries. Research shapes our products; engineering puts research to work. Offensive thinking helps enterprises build more dependable intelligent systems.
01Research-ledRESEARCH LED
02Evidence-firstEVIDENCE FIRST
03Always evolvingALWAYS EVOLVING
05WORK WITH US / ENGAGEMENT
Start with a real
security challenge.
From pre-launch AI assessments to continuous validation of live systems, we begin with your business environment and the risks that matter most.
Discuss your use case01What does an AI security assessment cover?
We assess LLM applications, RAG knowledge bases, AI agents and tool chains for prompt injection, sensitive-data exposure, permission-boundary failures and unintended execution. Application code and business workflows help us understand the full risk path.
02Can you work within our private environment?
We can discuss private-environment deployment and targeted assessments based on your data and infrastructure requirements. Scope, model access, data flows and deliverables are agreed before the project begins.
03What will we receive after the assessment?
Deliverables focus on reviewable findings: impact analysis, relevant paths, validation evidence and remediation guidance. Retesting and ongoing assessments can be arranged within the agreed scope.
YOUR NEXT MOVE.CHENGDU / OPEN TO COLLABORATION
Together, test the next boundary.
AI assessments / Product demos / Partnerships
Tell us what you are building. Let’s start with a concrete challenge.