Skip to content
VigilAI TRiSM
Sign inStart free trial

AI TRiSM Runtime Guardrail

Runtime trust for every AI agent you ship.

Vigil inspects every prompt, retrieved document and tool call — stopping prompt injection, data leaks and runaway agents in real time, and keeping the evidence your auditors ask for.

Start monitoring free Read the API docs

Free Starter plan includes 1 Agent, 1,000 logs/month, and 7-day retention (plus a 14-day full Team trial, no credit card required). Instrument in under 10 minutes.

Platform

One control plane for AI risk — from the first prompt to the board report.

  • Runtime guardrails

    Eight guardrail types — prompt injection, PII, secrets, toxicity, topics, tool allowlists, token budgets and custom patterns — evaluated on every call.

  • Live execution stream

    Watch every execution as it happens with verdicts, risk scores and latency. Pause, filter and drill into any trace in one click.

  • Immutable audit trail

    Context windows, retrieved documents and tool traces stored append-only, with credentials scrubbed before they ever touch disk.

  • Multi-tenant governance

    Separate organizations with Admin, Developer and Auditor roles, enforced at the database for every single request.

How it works

Two API calls around the agent you already have.

  1. 1

    Register the agent

    Add its endpoint, get an ingest key. Default guardrails attach automatically.

  2. 2

    Guard the input

    A pre-flight check blocks injection and redacts PII before the model sees it.

  3. 3

    Record the trace

    Send the output, context window and tool calls. Output guardrails decide what the user sees.

typescript
// Before the model: block hostile input
const guard = await vigil.guard({ input, retrieved_documents });
if (guard.decision === "block") return refuse();

// Call your agent exactly as you do today
const answer = await agent.run(guard.redacted_input ?? input);

// After the model: record the trace, enforce output guardrails
const verdict = await vigil.executions.create({ input, output: answer.text, tool_calls });
return verdict.decision === "block" ? withhold() : verdict.redacted_output ?? answer.text;

Threat coverage

Mapped to the OWASP Top 10 for LLM applications.

Deterministic detectors run in milliseconds at the edge, so guardrails never become your latency bottleneck.

  • LLM01

    Prompt injection

    Direct overrides, jailbreak personas, role-marker smuggling

  • LLM02

    Sensitive information disclosure

    PII redaction, payment cards, IBANs, SSNs

  • LLM05

    Improper output handling

    Secrets and credentials stopped in model output

  • LLM06

    Excessive agency

    Tool allowlists and an instant per-agent kill switch

  • LLM07

    System prompt leakage

    Extraction attempts detected and blocked

  • LLM08

    Vector & embedding weaknesses

    Indirect injection hidden in retrieved documents

Who Vigil is for

  • For CISOs

    A single risk posture across every AI system, with PDF evidence packs for audits and the board.

  • For security engineers

    Triage violations with full context — the prompt, the retrieved documents and every tool the agent touched.

  • For developers

    Ship faster with a test bench, a traffic simulator and guardrails you configure without redeploying.

See your agents the way an attacker does.

Load a sample workspace and watch real guardrails catch real attacks in under a minute.

Create your workspace