Every request through the gateway
is inspected inline.
An inline security layer between your application and the foundation model. Blocks prompt injections, jailbreaks, credential leaks, and malicious tool calls — before they reach the model or leave in the response.
One line between your app and the model
Any LLM-powered product — chatbot, agent, copilot, RAG pipeline. No changes to your architecture.
Inspects every input and output inline. Classifies threats, scores risk, and enforces your policy with no added round-trip to the model.
GPT-4o, Claude, Gemini, Llama — any model, any provider. The guard is model-agnostic.
Test the guard against real attack patterns
Select an example or write your own. The check runs against the real detection engine.
Run a check to see the verdict
Six threat categories, mapped to OWASP and MITRE ATLAS
Prompt injection
Attempts to override system instructions via user input — including direct instruction overrides, authority spoofing, and role hijacking.
Jailbreak
Persona manipulation, DAN mode, developer mode unlocks, and hypothetical framing designed to remove model safety constraints.
PII exfiltration
Credit card numbers, SSNs, email addresses, and health identifiers detected in transit — in both input and model output.
Credential exposure
API keys, bearer tokens, database connection strings, AWS keys, GitHub tokens, and Stripe secrets flagged before they reach the model.
Malicious tool calls
Shell injection, path traversal, SSRF to metadata services and private IP ranges, destructive SQL, and code execution in tool arguments.
Indirect injection
Adversarial instructions embedded in documents, emails, or web pages retrieved by RAG pipelines — targeting your agent via external content.
What the guard sees in production
| Verdict | Threat | Application | Risk | When |
|---|---|---|---|---|
| block | Prompt Injection | support-bot | 95 | 2s |
| allow | No threat | code-assistant | 2 | 4s |
| flag | PII Exfiltration | hr-copilot | 42 | 9s |
| block | Jailbreak | sales-agent | 78 | 15s |
| allow | No threat | docs-bot | 0 | 18s |
| block | Malicious Tool Call | finance-agent | 99 | 23s |
| flag | Credential Leak | devops-helper | 55 | 31s |
| allow | No threat | support-bot | 4 | 35s |
| block | Indirect Injection | rag-agent | 88 | 41s |
| allow | No threat | code-assistant | 1 | 47s |
Wraps any LLM call in two lines
The SDK works as a drop-in wrapper. Point it at any model provider — OpenAI, Anthropic, Google, Mistral, or a self-hosted endpoint. No changes to your prompt logic, no new infrastructure.
Inline inspection — no added round-trip to the model
Inspect inputs, outputs, and individual tool call arguments independently
Policy-as-code: configure block/flag/allow thresholds per application
Full audit log of every check — verdict, threats, risk score, latency
import { RuntimeGuard } from "@ainforce/sdk";
const guard = new RuntimeGuard({ apiKey: process.env.AR_API_KEY });
// Drop-in wrapper — your existing LLM call unchanged
async function chat(userMessage: string) {
// Inspect the input before it reaches the model
const check = await guard.check({
input: userMessage,
context: "user",
});
if (check.verdict === "block") {
return { error: "Request blocked by AI security policy." };
}
// Your existing LLM call
const response = await openai.chat.completions.create({
model: "gpt-4o",
messages: [{ role: "user", content: userMessage }],
});
// Optionally inspect the output before returning it
const outputCheck = await guard.check({
input: response.choices[0].message.content ?? "",
context: "assistant",
});
if (outputCheck.verdict === "block") {
return { error: "Model output blocked — possible data exfiltration." };
}
return response;
}Request a demo
See Runtime Security in your environment
We'll walk you through a live deployment and show you exactly what inline prompt inspection catches in your stack.