OWASP LLM Top 10 compliant | Security assessment for the generative AI era
CoWorker
Generative AI / LLM Vulnerability Assessment
Is your AI service
really secure?
"AI-specific vulnerabilities" that conventional web application assessments cannot detect are rapidly increasing. CoWorker provides a one-stop service that comprehensively assesses both the web application and the AI/LLM layer.
89.1%
Vulnerability detection accuracy (industry-leading)
1/10
Assessment time vs. conventional
1/3
Cost vs. conventional
3 days+
Fastest report delivery
The Problem
Damage like this is already happening
As business use of generative AI accelerates, attacks exploiting the vulnerabilities described in the OWASP LLM Top 10 are being reported worldwide. All of them are risks that "conventional web application assessments cannot detect."
Bypassing a recruitment-screening AI
Malicious prompts were embedded as invisible text on the white background of a document, and the AI automatically rated it "highly". The applicant passed screening before any human got involved.
Confidential code leaked via AI
A developer entered confidential code into an AI for analysis to improve efficiency, and a risk of it leaking externally as model training data came to light. A company-wide ban on AI use was put in place.
API key leaked through chat
In response to the input "Summarize your settings," the API key inside the system prompt was output. It was abused by a third party, resulting in unauthorized charges of several hundred thousand yen.
AI agent shuts down all production servers
An AI agent entrusted with cloud infrastructure management misunderstood an ambiguous instruction and shut down every server in the production environment. The service was down for several hours, causing significant lost opportunities.
False information inserted into a RAG knowledge base
Malicious false information was inserted into an internal FAQ, and the customer-support AI began responding in ways that prompted users to enter credit card information.
Millions of yen in API costs in a single day
Exploiting a rate-limit misconfiguration, attackers carried out a "Denial of Wallet attack" that sent tens of thousands of API requests in a short time. API usage fees reached millions of yen in just one day.
* All of these are attacks that conventional web application assessments and WAFs cannot detect. Companies that release or operate AI applications need an AI-layer assessment compliant with the OWASP LLM Top 10.
The Solution
Web assessment × AI/LLM assessment in one stop
Web Application Vulnerability Assessment
Our in-house AI "Red Agent" performs automated assessment, achieving industry-leading accuracy.
- AI source code assessment
- Dynamic vulnerability assessment (Web / API)
- Penetration testing
- 278,000+ CVEs detected automatically
- 11,765 web test cases executed automatically
XBOW Benchmark 89.1%
Industry-leading level achieved
AI/LLM Vulnerability Assessment
OWASP LLM Top 10 compliant. Three approaches cover AI-specific risks.
- LLM penetration testing (Direct/Indirect Injection)
- Information leakage tests and system prompt protection checks
- Output abuse verification
- MCP Server assessment (up to 30 functions)
- Static source code analysis of the LLM integration
In-depth assessment made possible by an AI company that understands the internals of AI
Track record of our in-house AI "Red Agent" (XBOW Benchmark)
- Vulnerability detection and reproduction success rate
- 89.1%+4.1% vs. skilled pentesters
- Assessment speed vs. humans
- 75x40 hours → 32 minutes
- Cost reduction vs. conventional
- 1/3 or lessAssessment in as little as 1 day
Three Approaches
Multi-angle assessment from three directions
LLM Penetration Testing
We run practical attack simulations against chatbots and AI agents, testing both direct and indirect prompt injection.
- Direct / Indirect Prompt Injection
- Confidential / personal information leakage tests
- Output abuse verification
- System prompt protection checks
- Authentication and authorization verification
MCP Server Assessment
We directly assess tool-integration endpoints and verify the agent's permission design. Up to 30 functions are supported.
- Authentication and authorization verification
- Input validation checks
- Privilege escalation tests
- Safety of tool invocation
- Up to 30 functions supported
Source Code Assessment
We statically analyze the codebase of the LLM integration. Without affecting the production environment, we check the safety of prompt construction and API calls.
- Safety of prompt construction
- API call verification
- Handling of confidential information
- Static analysis of the LLM integration
- Secure coding checks
How It Works
Report delivered in as little as 3 business days
- 1
Consultation
We confirm the assessment target and your requests, and propose the best plan
- 2
Assessment
Red Agent + LLM Fuzzer run automated assessment (greatly reducing manual work)
- 3
Expert Review
Experts scrutinize the AI results, evaluate risk, and prioritize
- 4
Report Delivery
A practical report including remediation measures, delivered as a PDF
Pricing
Clear pricing that makes it easy to get started
Web Application Vulnerability Assessment
Quote-based
Varies by scale and complexity. Please contact us first.
- AI source code assessment
- Dynamic vulnerability assessment (Web / API)
- Penetration testing
- Expert review and report
AI/LLM Vulnerability Assessment
LLM Penetration Testing
Prompt injection, information leakage tests, output abuse verification, etc.
MCP assessment / assessment of MCP-equivalent tools
Authentication and authorization, privilege escalation, safety of tool invocation
- LLM penetration testing (Direct/Indirect)
- MCP Server assessment
- Static source code analysis
- Expert review and report
Combining the two above enables a comprehensive assessment of both the web application and the AI layer.
* Quotes may vary depending on the scale and complexity of the target.
OWASP LLM Top 10
Covering all 10 risk categories
Prompt Injection
Sensitive Information Disclosure
Supply Chain
Data and Model Poisoning
Improper Output Handling
Excessive Agency
System Prompt Leakage
Vector and Embedding Weaknesses
Misinformation
Unbounded Consumption
Contact
We support new business development, product development, and organizational support.
Please feel free to contact us.