Skip to content
AUTONOMOUS PENTESTING

Your autonomous
AI hacker.

Hand it a target and an objective. It maps the attack surface, builds and runs the exploits, and writes up exactly what it proved. Launch from your terminal. Watch every move in the console.

Request beta access ↗Private beta · Authorized engagements
RECONNAISSANCEEXPLOITATIONREPORTING
[×] TARGET SYSTEM / 001Sample engagement · simulated output
ONE OBJECTIVE. EVERY ATTEMPT.

Give it a target.
Watch it work.

Launch from the terminal. Follow the attack as the agent maps the application, tests exploit paths, and proves execution.

INTERACTIVE DEMOSimulated assessment · Compressed playback
FROM FIRST REQUEST TO RCE

One click. Follow the complete attack.

SIMULATED TARGET · NO LIVE EXECUTION
[×] target.example / web assessmentREADY TO CONNECT
01 / TERMINALcrazypax — CLI
❯ crazypax login
Your terminal. One objective.
The agent handles the attack.
Signing in…
02 / ASSESSMENT GRAPH0 / 8
● Evidence path○ Unsuccessful attempt▧ Pending
LOGIN → SCOPE → EXECUTE
FOLLOW THE EVIDENCE

Every attack has a trail.

Select a terminal event or graph node to inspect the input, action, and result.

INPUT
Target + objective
RESULT
Waiting for launch
CAPABILITIES

What you can hand it.

A web pentest, a binary teardown, or both in one engagement.

Web applications & APIs

Endpoint enumeration, authn and authz testing, business logic, session security — then the exploit.

Reverse engineering & PWN

Disassembly, decompilation, debugging and exploit development against the binaries you hand it.

Forensics, cryptography & OSINT

Files, packet captures, broken crypto and investigations built from the artifacts and intel you supply.

PUBLIC WEB TARGETSARTIFACT UPLOADSMIXED ENGAGEMENTS
UNDER THE HOOD

One planner.
A crew of specialists.

The coordinator plans

Breaks the objective into bounded jobs, holds the evidence, and calls it when the goal is met.

Specialists execute

Dedicated executors for web and API testing, and for artifact work on binaries, captures and files.

One VM per assessment

Every engagement gets its own isolated VM. No model keys, no VMs, no tooling on your side.

STAY IN CONTROL

Limits it can't
talk its way past.

assessment runningSimulated run
Spending0% / 100%
budget set at launch
Active time0m / 60m
1h ceiling · default
Jobs0 / 20
20 ceiling · default

  1. 01

    Frozen scope

    Targets, authorization and acceptance criteria are locked before launch.

  2. 02

    Hard ceilings

    Spend, active time and job count are enforced by the platform — not by the model.

  3. 03

    Evidence-gated completion

    It's done only when reviewed evidence meets every criterion and the executor has stopped.

  4. 04

    Your controls

    Redirect it, stop it, reconnect later, export the report.

Built for AppSec and product security teams with more apps than testers, and for pentest firms that want one operator to cover more engagements.

OPERATOR CONSOLE

Every move.
On the record.

Watch tool execution, inspect the exploit evidence, read the report. Redirect the agent or kill execution from the operator controls.

Sample engagement · simulated output

[×] crazypax / sample-01 REPORT READY · SAMPLE
EVIDENCE / 03

Inspect the exploit evidence.

Verified in the sample

Remote code execution in document preview

The document renderer executes a controlled command. A second request returns a fresh proof marker and reproduces execution in this simulated assessment.

CONNECTED PATH / SAMPLE
  1. PUBLIC APP
  2. AUTH GATE
  3. SESSION
  4. RENDER API
  5. EXECUTION
  6. EVIDENCE
Evidence
POST /api/preview
Request artifact: preview-request.txt

HTTP 200 · renderer output
CRAZYPAX_PROOF_7F2A

Reproduction: fresh request / fresh marker
CRAZYPAX_PROOF_9B4C

Simulated evidence · payload omitted
Reproduction steps
  1. Identify the document preview endpoint and renderer input.
  2. Submit the controlled execution test and capture the returned proof marker.
  3. Repeat with a fresh marker and compare the request and response artifacts.
Suggested correction

Remove shell evaluation from the rendering path. Pass validated arguments directly to the renderer and restrict its process permissions.

Coverage limits

Execution demonstrated only in this simulated renderer. Persistence, lateral movement, and data exfiltration are not assessed.

SCOPE FIXED60 MIN / $10 LIMITNO LIVE EXECUTION
THE OPERATOR INTERFACE

Your terminal. Your engagement.

Detach and reconnect to the same engagement. Export its results when you need them.

Connect your CLI ↗
01 / AUTHENTICATE
$ crazypax login
02 / LAUNCH
$ crazypax assess
03 / RECONNECT
$ crazypax attach sample-01
04 / EXPORT
$ crazypax export sample-01 --output ./report
OPEN AND ENTERPRISE

Open source,
hosted for teams.

The Agent is planned as an open-source release, so you can audit the engine that tests your systems. Today the hosted platform runs it for you in isolated VMs — private beta, by invitation. Continuous assessments, compliance reports, integrations and private deployment are planned for teams.

YOUR NEXT ENGAGEMENT

Point it at
your attack surface.

One target. One objective. One autonomous AI hacker.

Request beta access ↗Private beta · Access by invitation
[×] CRAZYPAX / OFFENSIVE SECURITYILLUSTRATIVE ATTACK PATHS