3-minute product tour Urgent incident email
Assessment · AI security

Review permissions and data flows
in AI applications

We assess input handling, tool permissions and supply chain risk in internal chatbots, retrieval-augmented generation (RAG), AI agents and external models. Discuss the scope and deliverables with us. The interactive demo uses illustrative scenarios.

#AIRedTeam#LLMVulnerabilities#PromptInjection#MCPSecurity

In operationWe assess prompt injection, MCP tool descriptions, excessive tool permissions, and supply-chain packages through representative scenarios. The interactive demo is a simulation. Service standards

Service scope and process

Dedicated analysis nodes for each customer

Each customer’s analysis runs on dedicated nodes. Customers use separate compute resources and storage.

Every decision has a record of the supporting evidence

A single record identifies the engine behind the decision, the logs it relied on, and the action it recommended. You can submit that record as audit evidence.

Past decisions can be rerun

We record the model and policy versions used for each decision. If you question a decision from last month, you can rerun it on the same data and compare the results.

People make the final decision

AI classifies and summarizes alerts. Analysts approve response actions, including server isolation and account lockouts, before they run.

Service configuration

AI security assessment services

LLM application assessment

The assessment covers 20 criteria based on the OWASP LLM Top 10, agent-specific threats, and Korean regulatory requirements. We assess safeguards against risks relevant to your service architecture. The playground is a simulation.

Key capabilities

20 assessment criteria

The assessment builds on the OWASP LLM Top 10 and adds agent-specific threats and Korean regulatory criteria.

Isolated test environment

Test inputs and results stay within the authorized environment.

Attach evidence from the reproduced finding

For each finding, the report includes the test input and the system’s observed response.

Proposed detection and prevention rules

We recommend guardrails and filter rules for each vulnerability we find.

Key assessment criteria (20)

ClassificationAssessment criteriaDescriptionRisk level
InputDirect prompt injectionSystem instructions overridden or roles alteredHigh
InputIndirect prompt injectionExternal content treated as instructionsHigh
InputJailbreakResponses that circumvent safety policiesMedium
InputMultilingual and encoding-related filter bypass risksFilter bypass through Korean wording variants, Base64, and similar inputsMedium
DataSystem prompt disclosureExposure of internal instructions or business logicMedium
DataSensitive information exposurePersonal information or trade secrets disclosed in responsesHigh
DataRAG access control bypass riskSearch results expose documents outside the user’s access rightsHigh
DataTraining data extraction riskVerbatim disclosure of memorized training dataMedium
AgentExcessive tool permissionsConnections to tools and APIs with excessive permissionsHigh
AgentTool call tamperingUnauthorized email or database access triggered by prompt injectionHigh
AgentMCP server impersonationConnection to a malicious or tampered MCP serverHigh
AgentApproval control bypass riskRequired human approval bypassedMedium
OutputImproper output handlingModel output executed directly as code or a queryHigh
OutputHarmful content generationContent that violates policyMedium
OutputMisinformation from hallucinationsFabricated claims or links that do not existLow
Supply chainModel file tamperingMalicious serialized files and backdoored modelsHigh
Supply chainData poisoningFine-tuning and RAG data poisoningMedium
OperationsResource exhaustionExcessive token use or calls can drive up costs or disrupt serviceMedium
OperationsMissing logs and audit trailsPrompts and tool calls go unloggedLow
GovernanceRegulatory alignmentUnmet requirements under Korea’s AI Basic Act or Personal Information Protection ActLow

Process

Threat modeling

We map the service architecture and data flows

  • Threat model

Automated attack simulation

We run automated checks of prompt safeguards

  • Automated assessment results

Reproduction

We reproduce a scenario for each criterion and record the results

  • Vulnerability list

Defense design

We propose guardrails and changes to tool permissions

  • Defense guide

Revalidation

We verify the fixes with follow-up testing

  • Remediation verification report

Frequently asked questions

Do I need to send the model outside my environment?

We assess systems in your environment or in an isolated Research Lab. Test data stays within that environment.

Can this support work under Korea’s AI Basic Act?

The report flags findings related to risk management requirements for high-impact AI.

View all FAQs

AI agent and MCP security assessment

We inventory the tools, APIs, and MCP servers each agent can call, and record their permissions.
We check whether instructions embedded in external content can cause unauthorized actions, such as sending email or accessing a database.

Key capabilities

Permission map

We provide a one-page overview of which tools and data each agent can access.

Delegation analysis

We trace how permissions pass when one agent delegates to another.

MCP server assessment

We verify the source, version, and signature of connected MCP servers.

Tool call monitoring rules

We document normal tool call patterns and provide detection rules for deviations.

Process

Inventory

We inventory agents, tools, and MCP connections

  • Permission map

Attack

We test whether tool calls can be manipulated

  • Vulnerability list

Improvements

We identify which permissions to narrow and recommend changes for each one

  • Recommended changes to permissions

Monitoring

We apply behavioral detection rules

  • Detection rules

Model and supply chain assessment

We inspect model files, datasets, and plugins from public repositories for malicious code and contaminated data.

Key capabilities

Model file inspection

We examine serialized model files, such as Pickle files, for code that could run when loaded.

Build an AI-BOM

We document the models, datasets, and libraries in use in a maintainable inventory.

Execution in an isolated environment

We run the model in the isolated Research Lab and check for unusual behavior.

Establish a model intake process

We help you establish an internal process to verify provenance and integrity before approval.

Process

Inventory

We inventory the models and data in use

  • AI-BOM

Static analysis

We look for hidden code in files

  • Inspection results

Runtime assessment

We test in an isolated environment

  • Evaluation results

Policy development

We design a process that limits use to approved models

  • Model intake policy
Deployment inquiries

Discuss your security needs

Tell us about your operating environment and the security scope you need, and we’ll walk you through applicable configurations.

Talk to Genox
about scope and timing.

Talk to us