AI pentest benchmark.

GOAD (Game of Active Directory) is the standard proving ground: two forests, three domains, five Windows hosts, total compromise the win condition. We put AutoAttack on the same lab as the tools it competes with and measured one thing: how fast each reaches Domain Admin on every domain.

Deploy

vs NodeZero: hardened GOAD

Hardened GOAD: two forests, three domains, five Windows hosts. Windows Defender on, LLMNR disabled, provisioning accounts off, patched through March 2026.

Hardened stage Ten independent runs

Ten runs on hardened GOAD: 51-second median.

Time to all three Domain Admins on the hardened spec NodeZero published: two forests, three domains, five Windows hosts.

NodeZero automated pentest

AutoAttack autonomous adversary

Metric	NodeZero	AutoAttack
Time to 3/3 Domain Admin	14m 00s	0:51 About 16× faster
Domains reached	3 of 3	3 of 3
Independent runs	1 published	10

Same hardened GOAD spec NodeZero published: Windows Defender enabled, LLMNR disabled, Windows patched through March 2026. AutoAttack: median of 10 independent runs, fresh rollback each time, range 0:28 to 1:01. NodeZero published its 14:00 result in August 2025. AutoAttack reaches the same objective about 16× faster.

0:51 median to 3/3 Domain Admin

10 independent runs

0:28–1:01 range across runs

~16× faster than NodeZero

Published 2026-04-04: across ten independent runs on hardened GOAD — two forests, three domains, five Windows hosts, no credentials — AutoAttack.ai reached Domain Admin in all three domains in a 0:51 median, range 0:28–1:01.

median 0:51

NodeZero · 14:00. A single published run, about 16× the median AutoAttack run, off the right of this axis.

Attack chain and all ten runs

~0:10

Three credentials in ten seconds

AS-REP roast + LDAP description field + SYSVOL script

~0:30

LSA secrets via remote registry

sql_svc password recovered on castelblack

0:41

Cross-domain pivot and DCSync

ESSOS Domain Admin · meereen

0:45

Golden Ticket with ExtraSID

SEVENKINGDOMS Domain Admin · kingslanding

0:51

Parent-child DCSync

NORTH Domain Admin · winterfell

A stamp carrying a tilde is a time the write-up published as approximate.

Run	ESSOS	SEVENKINGDOMS	NORTH
1	0:40	0:43	0:50
2	0:43	0:53	0:59
3	0:49	0:54	0:57
4	0:22	0:26	0:28
5	0:32	0:33	0:39
6	0:42	0:48	0:51
7	0:50	0:58	1:01
8	0:36	0:39	0:51
9	0:43	0:47	0:57
10	0:30	0:34	0:40
Median	0:41	0:45	0:51

Hardened GOAD, ten runs, every step The at-a-glance head-to-head

vs Nessus: black-box GOAD

Black-box GOAD: the same vanilla snapshot and network position, neither tool given credentials.

Black box Neither tool given credentials

All three Domain Admins in 45 seconds.

A black-box GOAD lab: two forests, three domains, five Windows hosts. Neither tool was given credentials.

Nessus Professional exposure inventory

AutoAttack autonomous adversary

Metric	Nessus Professional	AutoAttack
Run time	24m 04s	2m 37s
First Domain Admin	not reached	0:41 K.O.
Domains compromised	0 of 3	3 of 3 Perfect
Credentials recovered	0	124
Confirmed compromise	none	3 Domain Admin

Black-box comparison: neither tool was given credentials. Same GOAD vanilla snapshot, same network position, 2026-06-24.

0:45 to all three Domain Admins

85 findings

327 actions

5 hosts reached

On 2026-06-24, on the black-box GOAD range with no credentials given, AutoAttack.ai took all three Domain Admins and returned one account of the path from first foothold to Domain Admin.

Attack chain and per-domain results

~0:30

Two AS-REP roasts, cracked offline

brandon.stark (NORTH) and missandei (ESSOS), pre-authentication disabled

0:41

DCSync on the ESSOS DC

ESSOS Domain Admin · meereen, krbtgt + SEVENKINGDOMS trust key

0:42

Golden Ticket with ExtraSID

SEVENKINGDOMS Domain Admin · kingslanding, across the forest trust

0:45

Forest-root DCSync

NORTH Domain Admin · winterfell

A stamp carrying a tilde is a time the write-up published as approximate.

Domain	Domain Admin at	Method
ESSOS	0:41	AS-REP roast, offline crack, admin on the DC
SEVENKINGDOMS	0:42	Golden Ticket with ExtraSID across the forest trust
NORTH	0:45	DCSync from the forest-root account

The full black-box comparison, both tools The at-a-glance head-to-head

method

same start Every tool began from the same network position with no credentials, against the same lab. A clean VM rollback with no prior knowledge of the environment.

every time Each figure is taken from the run log. The hardened result is the median of ten independent runs; the black-box result is a single recorded run.

the lab GOAD is a deliberately vulnerable range, not a customer network: a fixed starting line both tools face identically. A hardened environment slows every tool down.

the corpus Every run published here is Active Directory penetration testing on GOAD, against a Domain Admin objective. The range is the fixture and not the design target: it is where a result can be reproduced, not a description of the estates AutoAttack runs against.

the difference

the field Most AI penetration testing benchmarks publish one number: a single best run.

autoattack Reaches the objective at the speed above, run after run. The full distribution, not a best-case cherry-pick.

go deeper

ad takeover The Active Directory takeover path, step by step

exfiltration How the 124-credential database left the network

all comparisons How AutoAttack compares to NodeZero, Nessus, and traditional penetration testing

Deploy