[01] / AI Security Middleware

Stop prompt attacks before they reach your LLM.

The only guardrail system that thinks before it blocks.

Kataiq sits between your application and any LLM API and blocks prompt injection, jailbreaks, PII leakage, and toxic content using a parallel multi-agent detection pipeline.

Client
Kataiq
SA
EA
IA
CA
II
Aggregator
ALLOWBLOCKSUSPICIOUS
LLM Backend

Compatible with

OpenAIAnthropicOllamaAzure OpenAILiteLLM

[02] / The Threat Landscape

Public LLM apps are under attack

Prompt injection

Attackers smuggle instructions into user input and hijack your assistant's behavior.

Jailbreaks and policy bypass

DAN, developer mode, roleplay, and persuasion attacks defeat single-prompt guardrails.

Data leakage

System prompts, internal knowledge, and customer PII leak through generated responses.

[03] / Architecture

Multi-agent detection in parallel

Five specialist agents. One verdict. <500ms.

Client
Kataiq
SA
EA
IA
CA
II
Aggregator
ALLOWBLOCKSUSPICIOUS
LLM Backend

LangGraph fan-out / fan-in: every prompt is analyzed by specialist agents concurrently, not sequentially.

Each agent returns a confidence score and reasoning trail — no black-box verdicts.

Aggregator applies deployment policy thresholds and emits structured audit events for every decision.

[04] / Capabilities

Why Kataiq

Drop-in proxy

OpenAI-compatible endpoint, zero code changes.

Any LLM backend

OpenAI, Anthropic, Ollama, Azure, Bedrock.

Per-tenant policies

Different thresholds for different customers or environments.

Audit-grade telemetry

Structured JSON logs, run history, Prometheus metrics, exportable trails.

On-prem or SaaS

Docker Compose for self-hosted, hosted control plane on request.

Two-tier latency

80%+ of traffic resolved in <10ms by PreFilter. Full agent analysis in <500ms median.

[05] / Competitive Landscape

Why existing solutions fall short

Every current guardrail makes a single-shot decision. No reasoning. No depth. No adaptability.

SolutionMulti-AgentReasoning TrailParallel AnalysisOn-Prem
Lakera Guard
NeMo Guardrails
Prompt Armor
LLM-Guard
Kataiq

[06] / Technical Foundation

Built on proven security innovation

Three pillars of technical depth — not a weekend hack, not a wrapper around someone else's model.

01

Multi-agent AI orchestration

Production-proven patterns for coordinating specialized AI agents in parallel — battle-tested architecture for agent orchestration at scale.

02

Prompt injection domain expertise

Research-grade detectors developed through years of dedicated guardrails work — not a side project, a core competency.

03

Peer-reviewed AI security research

Published research in dual-path AI defense architectures in top-tier academic journals — theoretical rigor behind every design decision.

[07] / Integration

Architecture preview

Integration

# Before
client = OpenAI(api_key="sk-...")

# After — point at Kataiq instead
client = OpenAI(
    api_key="sk-...",
    base_url="https://proxy.kataiq.com/v1"
)
# Your application code is identical from here on.

Decision payload

{
  "verdict": "BLOCK",
  "confidence": 0.94,
  "attack_family": "prompt_injection",
  "agents": [
    {"agent": "semantic_injection", "score": 0.96},
    {"agent": "intent_classifier",  "score": 0.91}
  ],
  "request_id": "req_8h3d..."
}

[08] / Product Roadmap

Where we are. Where we're going.

v1.0Now

Multi-agent middleware. 6 of 7 phases complete.

v1.1Q3 2026

Performance optimization, red-team eval dataset, OpenTelemetry observability.

v2.0Q4 2026

ML fast-path: ONNX classifier handles 80%+ of traffic at sub-10ms.

v3.02027

Self-learning loop: novel attacks feed back to retrain the ML layer.

v4.02028+

Collective intelligence: network-wide threat learning across customers.

[09] / Get Started

Run Kataiq in front of your assistant in under an hour.

We're onboarding pilot partners now. Pilots include white-glove integration, custom policy tuning, and a co-authored security report at the end of the engagement.

We respond within 24 hours. No spam, no newsletter.

Built with v0