We're live on Product Hunt — support our launch and read reviews
Skip to content

The independent runtime control plane for AI agents

Stop prompt injections, data leaks & runaway agents — before they execute.

Real-time runtime guardrails for production AI agents. Enforce strict tool allowlists, block data leaks, and retain audit-ready execution traces. Your integration enforces the returned decisions.

No credit card required · Full Pro access for 14 days · Then free Starter forever

Vigiltrism AI - API-first AI guardrails to secure your LLMs in production | Product HuntFeatured on CodeHype
python
import os, requests

response = requests.post("https://vigiltrism.ai/api/public/v1/scan",
    headers={"Authorization": f"Bearer {os.environ['TRISM_API_KEY']}"},
    json={"input": user_message}, timeout=5)
if response.status_code != 200:
    raise RuntimeError("Vigil check failed")
guard = response.json()
if not isinstance(guard, dict) or guard.get("decision") not in ("allow", "flag"):
    raise RuntimeError("Execution halted")
# Never pass the raw message on: use the redacted version when Vigil returns one.
safe_input = guard.get("redacted_input") or user_message
if guard["decision"] == "flag":
    log_for_review(guard["violations"])
run_agent(safe_input)

Try it here — runs in your browser, nothing is sent

Try the Live Simulator
  • Deterministic Core

    Sub-millisecond pattern checks by default, with optional AI-assisted depth for advanced injection defense

  • OWASP mapped

    Checks addressing selected LLM risk categories

  • Audit-ready

    Workspace-scoped traces and evidence exports

  • HTTP API

    Pre-flight checks and execution trace ingestion

Built for security and platform teams shipping autonomous agents to production.

Works with your agent stack — one HTTP call, any framework

  • LangChain
  • LlamaIndex
  • CrewAI
  • AutoGen
  • Anthropic MCP
  • OpenAI Assistant
  • Custom REST / Webhook

Platform

Built for Production. Designed for Autonomous Agents.

  • Real-time runtime enforcement

    Evaluate prompts and retrieved documents before your model runs. Your integration must stop on a block or failed check and use redacted input when returned. Enabled guardrails depend on your plan.

  • Strict tool-call & agency control

    Check planned tool calls against allowlists before invoking them, apply token budgets and honor the per-agent kill switch. Recording a completed action cannot undo it.

  • Audit-ready compliance proof

    Structured, workspace-scoped execution traces with credentials scrubbed before storage and plan-based retention. Enterprise evidence exports support audits; they are not a certification.

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
// Integration outline — full HTTP examples in /docs:
// 1. POST /api/public/v1/scan with input and
//    retrieved_documents: [{ content: "Retrieved text" }].
// 2. Halt on block, HTTP errors or invalid decisions.
// 3. Give your model redacted_input when present.
// 4. Evaluate planned tools BEFORE invoking them.
// 5. POST /api/public/v1/executions with the trace.
// 6. Halt on block/errors; return redacted_output
//    when present. Never treat a failed scan as allow.

Threat coverage

Mapped to the OWASP Top 10 for LLM applications.

Selected risk categories are addressed by configurable detectors. Coverage depends on enabled policies and your integration; no detector catches every attack.

  • 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.

Measured, not marketed

Protect your AI agents in minutes

On our published 1,000-sample benchmark, Vigil with the optional AI check caught 47.0% of attacks at a 1.0% false-positive rate — versus 23.5% for Prompt Guard 86M on 400 samples. See the method

No credit card required · Full Pro access for 14 days · Then free Starter forever

  • 47.0%

    attacks caught (AI check on)

  • 1.0%

    false positives

  • ~0.05 ms

    core check CPU time

  • 0

    prompts used for training

FAQ

Questions CISOs and developers ask first.

Contact

Questions, billing, privacy requests or a security report? Email support@vigiltrism.ai.