agent red

autonomous offensive security

hack yourself before someone else does.

agent red puts a crew of AI operators inside your own environment. They attack what you own around the clock — and only tell you about what they actually got into, scored and ready for your auditor.

Start freeWatch it work
ssh agent-red-loopy-doo · over your tailnet
watchsurface monitoringclaude-sonnet-4.5

❯watch everything acme.example.com exposes. tell me when it changes.

●Sweeping every six hours and diffing against the last inventory. I'll only speak up when something moves.

└ enumerate subdomains · 1,909 names[ran unattended]

└ probe live hosts · 412 up[ran unattended]

●New since Tuesday: grafana-stg.acme.example.com is up on :3000 and answering without a login.

high · CVSS 8.6dashboards readable by anyone on the internet

└ file intel · deduped against 1,204 prior[ran unattended]

●Filed, scored, mapped to CWE-306. Back to the loop — nothing else moved.

acme.example.com · unattended 9d · 412 hosts tracked · ctx 14% · $2.18 today

scanner · tag:scanner

subfinder  acme.example.com    1,909 names
dnsx       resolve               412 live
naabu      top-1000               71 open
httpx      probe                 188 http
nuclei     templates=4,412         1 match

[high] grafana-stg.acme.example.com:3000
       exposed-panels/grafana-unauth
       anonymous org role = Viewer

next sweep in 05h 41m · nobody attached
Three operations, one agent, running on a box you own. Click the rail — it's the real layout.
operators on shift
24/7
per action
$0.001
pricing tiers
0
inbound ports opened
0

why agent red

Eight things nobody else does all at once.

  1. 01

    Hack yourself before someone else does.

    Attackers don't wait for your annual pentest. agent red runs the same playbook they do — recon, foothold, escalation — against systems you own, and only reports what it actually got into.

    surface monitoringevery 6h
    penetration teston demand
    adversary emulationgoal-seeking
    ai red teamprompt → tool → data
  2. 02

    Set it and forget it.

    Operators don't clock off. Close the laptop, go on holiday — they keep sweeping, diffing and probing around the clock, and only interrupt you when something moved or they need a yes.

    quiet archive● unattended 9d
    long winter● on every push
    paper tiger◉ needs you
    youasleep
  3. 03

    Pay for what you use. That's the whole price list.

    Two meters: AI tokens at the provider's rate, and actions at $1 per 1,000. No seats, no per-target fees, no plans. Every capability is open to everyone — enterprises just buy actions in bulk, cheaper.

    ai tokensat cost
    actions$0.001 each
    per target / seat$0
    tiersnone
  4. 04

    Your environment. Your rules. Your hands on the wheel.

    Operators run in a container on infrastructure you own — never ours. Let them assist while you drive, take over the terminal mid-run, or set them loose and read the intel later. You choose how walled-off they are.

    assistyou drive, it suggests
    take overgrab the tui mid-turn
    set & forgetruns unattended
    runs onyour box
  5. 05

    Intel your auditor already speaks.

    Every finding arrives scored in CVSS v4.0, tagged to CWE and MITRE ATT&CK / ATLAS, with reproduction steps, remediation and hashed evidence. Drop it straight into your SOC 2, ISO 27001 or PCI evidence trail.

    severityhigh · CVSS 8.6
    cweCWE-306
    att&ckT1190
    evidencesha256:9f2c…e41a
  6. 06

    Reach your operators from anywhere.

    Every agent joins a private tailnet that only your org can see. One line from any laptop and you're SSHed in — no open ports, no VPN to run. Remove someone from the org and their devices drop off instantly.

    $ tailscale up --authkey ••••

    $ ssh agent-red-loopy-doo

    inbound ports0
    reachyour org only
  7. 07

    A team, not a scanner.

    A handler plans the operation and runs a crew of operators across your agents. They leave each other notes, pick up where one left off, and survive restarts — the operation resumes, it doesn't start over.

    handlerplanning
    └ recon● working
    └ breach◉ blocked
    └ report○ idle
  8. 08

    It can't grade its own homework.

    The AI proposes a CVSS vector; our gateway re-scores it, so severity can't be inflated. No provider keys ever sit on your box. Nothing is attacked until you prove you own it. And a stealth mode keeps it from tripping your own SOC.

    severityre-scored server-side
    keys on agent0
    ownership proofdns txt ✓
    egressstealthy

the difference

most tools hand you 400 maybes. this one shows you what actually broke.

Scanners flag anything that looks suspicious and leave you to sort it. agent red tries the weak spot, chains what it finds, and reports it only once it works. If it says it's real, it got in.

$0per app, server or repo

Most security tools charge per app, per seat, per tier — so most teams test a slice of what they own with half the features. We don't. Every capability is open to everyone, and you pay only for the tokens and actions your operators actually use.

How billing works

only test what you own

Put an operator on shift tonight.

Start with free credit, point it at something you own, and read what it found in the morning.

Start freeSee the pricing