Skip to content
offense · research · defense

Break it.
Prove it.
Lock it down.

An AI agent that attacks your own apps. A hunter that checks whether the patch actually closed the hole. A daemon that bans what hits your servers. One rule across all three: no proof, no finding.

Agent · private beta0day hunter · PHP results publishedWarden · open source soon
crazypax · opssample output
agentsimulated
$ crazypax assess
scope frozen · 1 target · authorization on file
ceilings spend ✓ time 1h ✓ jobs 20 ✓
[recon] 42 endpoints mapped
[web] session handling → candidate
[exploit] proof executed in scope
[evidence] request, response, screenshot saved
finding F-01 · high · evidence reviewed
0day hunterillustrative
advisory ADV-SAMPLE-001 · fix landed
- if (refcount == 1) free(obj);
+ if (refcount == 1 && !reentrant) free(obj);
? sibling path: destructor re-entry
✓ sanitizer heap-use-after-free
✓ reproducer id → uid=33
✗ patched build · ✗ gadget removed
tier → CONFIG-GATED
wardensimulated
sshd Failed password for root from 203.0.113.24
nginx 203.0.113.24 GET /.env 404
nginx 203.0.113.24 GET /?f=../../etc/passwd
detect web.traversal + scanner → incident #118
propose ban 203.0.113.24 · 24h
policy auto · never-touch clear
enforce nftables ban · expires 24h

Three tools.
Zero hand-waving.

Offense, research and defense, built on the same rule: a model can suggest anything, but only evidence and deterministic code get the final word.

Crazypax AgentPrivate beta

Your autonomous AI hacker.

Hand it an authorized target and an objective. It maps the attack surface, builds and runs exploits, and writes up exactly what it proved.

Meet the Agent →
Crazypax 0day hunterService · PHP first

Patched ≠ fixed.

An advisory says the bug is gone. The hunter checks whether a sibling path survived and whether your stack can reach it. No reproducer, no finding.

Meet the 0day hunter →
Crazypax WardenOpen source soon

The model talks. The policy decides.

Host defense for Linux. Reads your logs, correlates attacks into incidents, bans with expiring nftables rules. An LLM can explain. Only policy can block.

Meet Warden →
how they fit

From advisory
to ban.

How a team runs the three today. Each tool works on its own and the handoffs are manual — wiring them together is on the roadmap, and marked that way.

01 0day hunter

Is it reachable?

An advisory drops. The hunter diffs the fix, hunts the sibling path the patch missed, and proves it under a sanitizer with a reproducer — or tells you exactly why it doesn't apply to your stack.

Service · PHP first
hunter · verdict record
advisory ADV-SAMPLE-001 · use-after-free fixed
- if (refcount == 1) free(obj);
+ if (refcount == 1 && !reentrant) free(obj);
? sibling path destructor re-entry
✓ sanitizer heap-use-after-free
✓ reproducer id → uid=33
✗ patched build negative
✗ gadget removed negative
Tier · CONFIG-GATED
Illustrative record · loopback lab
02 Agent

Can they get in?

Point the Agent at your own apps inside a frozen scope. It maps the surface, builds and runs the exploit, and comes back with evidence — not a list of maybes.

Private beta
agent · engagement
$ crazypax assess
scope frozen · 1 target · authorization on file
[recon] 42 endpoints mapped
[web] session handling → candidate
[exploit] proof executed in scope
[evidence] request, response, screenshot saved
finding F-01 · high · evidence reviewed
Evidence · reviewed
Sample engagement · simulated output
03 Warden

Shut the door.

Warden tails your logs, correlates the noise into one incident and bans the source with an expiring nftables rule — under your policy, never on a model's say-so.

Open source soon
warden · incident
sshd Failed password for root from 203.0.113.24
nginx 203.0.113.24 GET /.env 404
nginx 203.0.113.24 GET /?f=../../etc/passwd 400
detect web.traversal + scanner → incident #118
propose ban 203.0.113.24 for 24h
policy auto (rule: web-traversal) · never-touch clear
enforce nftables ban · expires in 24h
Banned · expires 24h
Illustrative incident · simulated output
04 Feed back

Close the loop.

Hunter detection packs delivered to Warden as a rules feed. Agent findings turned into Warden watchlists. On the roadmap. Not shipped — and we won't pretend otherwise.

Planned
roadmap · not shipped
hunter detection packs ┄┄▶ warden rules feed
agent findings ┄┄▶ warden watchlists
shared evidence records ┄┄▶ one console
status: planned · today a team runs the handoffs
Planned
Planned integration · not available today
1VMisolated VM per Agent assessment. You never touch keys, VMs or tooling.
4.1Mfuzz executions — published as a negative result.
2,500+ CVEscovered by Warden's ~3,000 sensitive-path probes.
0bans on a model's say-so. The LLM advises; Warden's policy decides.
house rules

Non-negotiable.

  1. Scope is frozen.

    Targets, authorization and limits are locked before anything runs. No tool can widen them mid-flight.

  2. Proof or it didn't happen.

    A crash isn't an exploit. A scan isn't a finding. Every result ships with reproducers, receipts and stated limits.

  3. Code holds the leash.

    Models propose. Deterministic code enforces scope, budgets and policy — and you approve what matters.

open core

Read the code
that tests you.

All three products are going open source — read, run and audit the code that tests and defends your systems. Teams pay for hosted operation, continuous coverage, fleets, feeds and compliance. Safety controls are never a paid feature.

built for

Teams running software
they didn't write.

Hosting and platform providers, agencies and MSSPs carry thousands of third-party components across many apps and servers. Start by verifying which patch gaps your stack can actually reach, hit your apps with the Agent, and defend the fleet with Warden.

your move

Point it at something you own.

Private beta · Authorized engagements only