[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.
Compatible with
[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.
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.
| Solution | Multi-Agent | Reasoning Trail | Parallel Analysis | On-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.
Multi-agent AI orchestration
Production-proven patterns for coordinating specialized AI agents in parallel — battle-tested architecture for agent orchestration at scale.
Prompt injection domain expertise
Research-grade detectors developed through years of dedicated guardrails work — not a side project, a core competency.
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.
Multi-agent middleware. 6 of 7 phases complete.
Performance optimization, red-team eval dataset, OpenTelemetry observability.
ML fast-path: ONNX classifier handles 80%+ of traffic at sub-10ms.
Self-learning loop: novel attacks feed back to retrain the ML layer.
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.