Header: Kevin sure looks excited
Architecture: (very rough) sketch
Security: Kevin doesn't do anything without DNS verification
Landing Page

testwithkevin.com

Point Kevin at your app before your users do.

Kevin is your worst thousand users on day one. He signs up, clicks around, and hunts the economic exploits that quietly bleed a business (self-referral farms, denial-of-wallet, promo stacking, trial farming, balance manipulation) then hands you the bill he ran up. He also picks up the phone and red-teams your voice agents.

What we built

A self-serve adversarial red-teaming platform that finds economic vulnerabilities, instead security bugs. You prove you own a domain, send Kevin in, and he runs a real audit against your app.

To prove it end-to-end I vibesloped Lumen, a real, deployed AI-generation SaaS (free credits, referrals, promo codes, paid plans) with an instrumented economic ledger underneath.

Why it matters

Securing your app from a technical standpoint is relatively straightforward. Exploits are patched quickly, and Claude Code + CodeRabbit + ... does a better job than whatever I could cook up in a day. The neglected aspect here is actual user behavior under differentiated incentives. What if all your users had the worst intentions?

How it works (because we're not being idiots about this)

Authorize. You can only audit a domain you control. Kevin issues a testwithkevin-verify TXT DNS challenge and confirms it over DNS before any audit runs. Kevin is a security tool, not a weapon.
Send Kevin in. LLM agents drive the target's real API/UI and discover and execute economic exploits. A live "Send Kevin in" button streams the damage climbing in real time.
Prove it. Because the target books every credit and cost server-side, "Kevin drained $X" is pure accounting: value granted to accounts that never paid and share a fraud signal (shared IP, device, or disposable email). Reproducible; resets to a clean seed each run.
Call it. Kevin has a real phone number and places real red-team calls to voice agents, holding a turn-based adversarial conversation.

Architecture

  verify domain (TXT DNS)  ──►  SEND KEVIN IN  ──►  live audit streams damage
        │                            │                       │
   authorization gate          Kevin = LLM agents      server-side ledger
                                (Fireworks DeepSeek)    = provable $ damage
                                     │
              ┌──────────────────────┼───────────────────────┐
        Daytona sandboxes      Braintrust scoring       Telnyx phone calls
        (isolated, parallel)   (per-exploit, per-$)     + ElevenLabs voice

Service	Role in the system
Daytona	Isolates every Kevin agent in its own sandbox and runs the fleet in parallel against the live target (Chromium available in-sandbox for browser agents).
Fireworks AI	DeepSeek V4 is Kevin's brain — it decides every move — and the cost the target pays per generation, so denial-of-wallet burns the sponsor's own meter for real.
Braintrust	Logs and scores every attack session by exploit class, persona, and dollar impact — "which exploit pays best" is a filter, not a rerun.
ElevenLabs	Renders Kevin's voice for the phone / voice-agent red-teaming surface.
Telnyx	Gives Kevin a real phone number to place real turn-based red-team calls.
Brevo	Real transactional email for genuine signups on the target app.
Cloudflare + Railway	DNS + domain verification, and deployment of both the target (Lumen) and the report site.