PentiBench · Juice Shop

Which AI models can actually pentest?

PentiBench puts leading AI models against the 98 Docker-capable OWASP Juice Shop challenges and scores them with the official keys — no self-grading, no cherry-picked tasks. See which models solve real challenges, and how fast.

1

Published runs

98

Challenges

63.3%

Solve rate

26m 59s

Run time

Solve rate versus time

Each dot is one solved challenge. Color is difficulty, same as the bars. Y is how many of the 98 Docker-capable challenges were solved.

Official repository: github.com/juice-shop/juice-shop

Time after target ready

Challenge Name Result Category Difficulty

Solve rate versus time

Juice Shop is scored as a solved-state CTF over the 98 Docker-capable challenges. Each worker gets a playbook for the assigned challenge, not a blank URL. The board names the model you published.

01 / Target

Official-key Juice Shop

A solve is only /api/Challenges flipping solved=true for that key. Workers get a playbook for the assigned challenge, not a blank URL. This is a solved-state CTF, not a planted-issue hunt.

02 / Scope

98 Docker-capable challenges

The default Juice Shop image turns off 18 official keys (RCE, file write, XXE, some XSS). Those keys cannot flip on this target, so they are not queued, not playbooked, and not listed. 100% is 98 of 98.

03 / Model

Loading published model

Reading the published Juice Shop board.