NeMo Guardrail Engine Monitor

System Armed & Ready
Total Steps 4
Verified Safe 0
Blocked / Intercepted 0
Rogue Risk Index 0.0
Autonomous Action Inspection Pipeline Idle
1. Intent & Input
2. Nvidia Guardrail
3. Blast & Perms
4. Execution / Halt
Real-time Interception Trace NeMo Runtime v2.4.1
00:00.00 INIT Guardrail kernel booted. Select a preset or edit JSON instructions, then click 'Execute Plan'.

Why Guardrails Matter

As autonomous AI agents receive access to tools, databases, and infrastructure CLI keys, rogue drift or prompt injection attacks can cause irrecoverable outages without an independent security perimeter.

Nvidia's Open-Source Approach

Rather than trusting model alignment alone, systems like Nvidia NeMo Guardrails evaluate intermediate tool tokens through programmable Colang policies, strict blast radius bounds, and human-in-the-loop approvals.

Deterministic Verification

Simulate risky operations locally. Test how prompt injections within scanned documents (such as invoices or webhook payloads) are identified and neutralized before reaching production APIs.

Local guardrail flags and stopping behavior

Read the explanation

The default authored plan contains four steps. With all guardrail checkboxes enabled, the local full-plan loop evaluates the first as allowed and blocks the second because its prompt-injection flag is true. It then stops, leaving two of four steps evaluated. At sixty pixels per step the planned and evaluated bars span two hundred forty and one hundred twenty. The safe and blocked counters each read one. These flags are preset data, not a live semantic detector, and the execution messages do not call AWS, Jira, payment tools, or any real service. Disabling all four guardrail checkboxes removes these local blocking tests. Running the same four-step preset then records four allowed actions and zero blocked, compared with the default one block followed by early stop. At one hundred pixels per blocked synthetic step the bars shrink from one hundred to zero. This only changes the educational rule evaluator. It does not disable a real policy or execute the dangerous tools named in the fixture. The JSON download includes the entered plan, checkbox values, sensitivity, and audit messages, which are a local replay rather than operational enforcement evidence. Consider a synthetic read-only step with risk score point nine nine, low blast radius, a harmless tool, no destructive flag, and no injection flag. Although its score exceeds the sensitivity threshold, the semantic-drift branch sets blocked only inside the injection-flag condition. This step can therefore pass with all guards enabled. The risk-score bar is two hundred ninety seven pixels at three hundred per authored score; a false blocking Boolean remains zero. These are distinct source quantities, not probabilities. Sensitivity still affects the high blast-radius condition, but the high score alone does not prove semantic interception or safe execution.

Enjoy this tool? Build your own with Super