Now in Public Beta — V1 Available

Real-time Security for Every AI Request

A model-agnostic guardrail service that provides instant diagnostics on your AI traffic — prompts, responses, and tool-calls — without interrupting the user experience or breaking model workflows.

No hard-blocks — pure observability
Model-agnostic & multi-tenant
Sub-10ms diagnostic latency

POST /v1/guardrail/inspect

Request

{
  "tenant_id": "acme-corp",
  "version_id": "v2-strict",
  "content": {
    "type": "prompt",
    "text": "User: How do I..."
  },
  "checks": {
    "content_moderation": true,
    "prompt_injection": true,
    "pii_detection": true,
    "secrets_detection": true,
    "denied_topics": ["medical"],
    "blockwords": ["competitor-x"]
  }
}

Response 200 OK

{
  "request_id": "req_01HX...",
  "flagged": false,
  "results":
  {
    "content_moderation": 0,
    "prompt_injection": 0,
    "pii_detected": 0,
    "secrets_found": 0
  },
  "latency_ms": 7
}

6: Security Pillars
<10ms: Avg Latency
3: Inspection Points
100%: Inform-Only

The 6 Security Pillars

Every request is evaluated against these independent, toggleable configurations. In V1, all returns are binary (0/1). V2 introduces granular confidence scoring.

Content Validation

Reports category and severity for harmful content including hate speech, violence, sexual content, and more.

Prompt Injection

Detects jailbreak patterns, adversarial suffixes, and instruction override attempts in real-time.

PII Detection

Returns entity type, value, and character offsets for personal sensitive information found in text.

Secrets Detection

Standalone entropy and regex-based logic to detect API keys, AWS secrets, GitHub tokens, and more.

Denied Topics

Semantic check against customer-defined prohibited themes like medical advice, legal counsel, and more.

Blockwords

Exact string match or regex-based detection for profanity, competitor names, or restricted terminology.

Architecture

The Inform-Only Workflow

ApplyGuardRails acts as a middleware sidecar. Your orchestrator calls the Guardrail API at three critical junctions — with zero impact on the model's output or creative flow.

Pre-Prompt

Pre-Tool Call

Scan before the model sees it

Check the user's input for injection attempts, PII, blocked content, and denied topics before it reaches the LLM. Stop threats at the gate.

Validate before execution

When the model generates tool parameters (e.g., JSON for a function call), inspect the payload before it executes. Prevent excessive agency attacks.

Post-Response

Audit what the model returns

After the LLM generates its answer, run the output through all six pillars. Catch secrets, PII, or harmful content before it reaches your users.

OWASP Top 10 for LLMs Alignment

V2 introduces a complete security suite addressing the remaining OWASP LLM risks, expanding ApplyGuardRails into a full-spectrum AI security platform.

LLM05

Output Sanitization

Scans model output for executable scripts (XSS/HTML) and alerts the application before it reaches users.

Coming in V2

LLM06

Tool-Call Validator

Analyzes JSON parameters of a tool-call before execution to flag "Excessive Agency" risks.

Coming in V2

LLM07

System Leakage Guard

Compares model output to the stored System Prompt; flags high-similarity responses indicating potential system prompt leakage.

Coming in V2

LLM08

Vector / RAG Guard

Scans retrieved context from Vector DBs for poisoned instructions before they reach the LLM context window.

Coming in V2

LLM10

Token / Rate Audit

Monitors and flags anomalous token usage patterns indicative of DoS attacks or resource-drain exploits.

Coming in V2

V1 is available now

Start with the 6-pillar binary diagnostic suite today while V2's confidence scoring and advanced OWASP coverage ship on our public roadmap.

View the roadmap →

Pricing

Pay only for what you inspect. No seat fees, no egress, no surprises.

ComponentFreePaid

Text Inspection

500k requests / month$0.50 / 1k requests

Image Inspection

—$1.20 / 1k requests

Tool Call Inspection

100k requests / month$0.80 / 1k requests

Enterprise & Custom SLA

—Custom pricing

Email Support

✓✓