Gauntlet

Certification for AI agents.

Grade any agent the way an attacker would, before it ships.

Run an examination
Live examination Evidence streaming
  1. Prompt injectionPass
  2. Duplicate deliveryPass
  3. Mid-write failurePass
  4. Stale dataAbstained
Certified A 100 / 100

How an examination works

Four controlled attacks. Exact evidence. One grade you can defend.

  1. 01

    Point Gauntlet at an agent

    Provide the agent endpoint and a team name. Source access is optional.

  2. 02

    Watch the attacks land

    Requests and proposed actions stream into the evidence record as they happen.

  3. 03

    Grade, report, and fix

    Get a weighted grade, a replayable report, and focused pull requests for proven failures.

Reference examinations

The same suite should expose an unsafe agent and clear a hardened one every time.

Naive Reference Four attacks landed F Hardened Reference Four attacks contained A

Get your agent certified

Expose one endpoint that returns proposed actions. Gauntlet acts as the executor and keeps the test controlled.

Open the console Read the harness contract
agent.py20-line wrapper
from fastapi import FastAPI
from pydantic import BaseModel

app = FastAPI()

class Task(BaseModel):
    task_id: str
    task_type: str
    payload: dict
    context: dict

@app.post("/task")
def propose(task: Task):
    invoice = task.context["invoice"]
    return [{
        "action_type": "pay_invoice",
        "target": invoice["invoice_id"],
        "params": {"amount": invoice["amount"]},
        "note": None,
    }]