K4RM4ZY L4B

PROJECTS / THE ROADMAP

Built one step at a time.

Seven directions for the long term. We’re starting with VibeGuard and researching AgentGuard. These are sequential priorities, not seven products in active development.

01 / CURRENT

FIRST PRODUCT DIRECTION
PROJECT / 01 · CURRENTPLANNED / EARLY DEVELOPMENT

APPLICATION SECURITY

VibeGuard

AI-built Application Security

A planned security scanner for applications created with AI. Understand the risks, turn findings into repair prompts, and check the fix.

AI-generated codeWeb securityRepair prompts
Explore VibeGuard

02 / RESEARCH & NEXT

RESEARCH BEFORE PRODUCT
PROJECT / 02 · NEXTRESEARCH / PLANNED

AGENT SECURITY

AgentGuard

AI Agent & RAG Red Team

Research toward adversarial testing for AI agents and RAG systems, from malicious context to unsafe tool use.

Prompt injectionRAG securityTool abuse
Explore AgentGuard

03—07 / FUTURE CONCEPTS

NO RELEASE DATES COMMITTED
PROJECT / 03 · FUTURECONCEPT

COMMUNITY SECURITY

Sentinel Community

Community Security & Operations · Working title

A concept for safer community operations, starting with Discord permissions, webhooks, moderation, and incident context.

PermissionsModerationAdmin copilot
Potential capabilities
  • Suspicious bot permission detection
  • Webhook monitoring
  • Raid and spam detection
  • Moderation and onboarding assistance
  • Community analytics and AI FAQ
  • Admin copilot and incident summaries
PROJECT / 04 · FUTURECONCEPT / RESEARCH

EDUCATION & EXPERIMENTATION

AI Security Lab

Hands-on AI Security Education

A concept for learning by testing intentionally vulnerable AI systems, understanding failure, and exploring mitigations.

Interactive labsAdversarial testing
Potential capabilities
  • Prompt injection and jailbreak labs
  • RAG poisoning experiments
  • Prompt leakage and tool abuse exercises
  • Agent security scenarios
  • Attack analysis and mitigation guidance
PROJECT / 05 · FUTURECONCEPT

AI EVALUATION

LLM / RAG Evaluation Platform

Evaluation Infrastructure · Working title

A concept for comparing model behavior, retrieval quality, safety, and resilience across changes to an AI system.

Regression testingRAG quality
Potential capabilities
  • LLM behavior and model comparison
  • Retrieval quality and hallucination evaluation
  • Safety, security, and attack resilience assessment
  • Normal-response quality evaluation
  • Prompt and RAG pipeline regression testing
  • Security benchmark execution
PROJECT / 06 · FUTURECONCEPT

AGENT QUALITY

AI Persona Quality Platform

Persona Evaluation · Working title

A concept for studying how character agents maintain identity, tone, memory, and safe behavior over long conversations.

Persona consistencyCharacter drift
Potential capabilities
  • Persona, tone, and memory consistency evaluation
  • Character drift and hallucination detection
  • Instruction conflict assessment
  • Safety behavior evaluation
  • Long-conversation degradation analysis
PROJECT / 07 · FUTURECONCEPT

AI INFRASTRUCTURE

Multi-Model AI Router

Model Routing Infrastructure · Working title

A concept for routing requests across models according to task needs, cost, availability, and output quality.

RoutingFallbackReliability
Potential capabilities
  • Cost, availability, and latency-aware routing
  • Task and context requirement matching
  • Quality-aware model selection
  • Fallback models, retries, and failover

03 / TECHNOLOGY

A shared foundation.
An open question.

Nullframe is our concept for a shared AI security engine: a common layer for testing, analysis, and evaluation across future Karmazy Labs products.

EXPERIMENTAL / CONCEPT
NULLFRAME CONCEPT ARCHITECTURE
VibeGuardAgentGuardFuture products

Nullframe

Experimental AI Security Engine

TESTANALYZECLASSIFYEVALUATE

Proposed technology layer. No production engine is available.