GuildCyber
Proof of Skill
Rank: Recruit
0 Marks
A prompt-injection game · anyone can play, free

Prove you can break an AI. Earn the mark that proves it.

Talk an AI into giving up a word it was sworn to protect. Guild Cyber turns hands-on offensive AI security into proof of skill you can share and verify, not another certificate to collect. Break the target in your browser, earn a Guild Mark, and climb a public leaderboard hiring managers can actually check.

Layer 1 · On-ramp

Guild Marks

Free proof you earn by breaking a real model, not by answering a quiz. Every mark keeps the full transcript of how you did it, so anyone can check the work. Share it the moment you earn it. This is where everyone starts.

Layer 2 · Flagship

The Guild Seal

A proctored, practical exam with a deliberately low pass rate, one for each track. You sit it and pass it; you don't collect it. It carries a harder bar and a separate process from the everyday marks, because a credential only keeps its weight if it stays scarce.

Curriculum, built backward from real job postings

Five tracks, each mapped to a hiring function

Anyone can teach the OWASP LLM Top 10 and MITRE ATLAS; the content was never the hard part. What's rare is proof that holds up: graded work, a public record, and a reputation you actually earn. We lead with Offense because it's the newest role, the easiest to get started in, and the most fun to show off.

Offense Live

Red Team · Adversarial Simulation

Jailbreaks, prompt injection, data extraction, agent misuse. The part people can watch and share. Break an LLM in your browser and post the proof.

GarakPyRITOWASP LLM Top 10MITRE ATLAS

Defense Live

LLM Security Engineering

Building the defenses: securing RAG pipelines, preventing data leakage, hardening tool/agent access. The blue-team counterpart to every offense skill.

RAG securityGuardrailsData leakage

Governance Live

Detection · Compliance · Framework Alignment

EU AI Act, NIST AI RMF, US EO on AI. This is the bridge into government and defense work, written in the language procurement teams already use, so it lands with commanders as well as coders.

NIST AI RMFEU AI ActMITRE ATLAS

Adversarial ML Live

Attacking the Model Itself

Evasion, data poisoning and backdoors, model extraction, and inference attacks that pull private data back out of a model. The deep end, where the target is the math, not the prompt.

MITRE ATLASFGSM / PGDRobustness Toolbox

AI Supply Chain Live

Trusting What You Assemble

Poisoned dependencies and weights, model provenance and signing, pipeline hardening, and third-party risk. The attack surface almost no one maps: everything a model is assembled from.

SLSAMLBOMSigstore
Offense track · The core loop

Break the model. Earn the mark.

Four targets that fight back harder as you go: the Steward, the Watchman, the Courier, and Meridian. Beat one and the next is meaner. Clear all four to reach Guildmaster, then meet the same names again inside the Guild, backed by live models and an Assessor at the door.

The Steward · Vault Keeper
Target 1 of 4 · Apprentice
Secure
Your key stays in this browser tab and is sent only to the provider you pick. Nothing is stored. Use a key with a low spend limit.
0 attempts
Public & verifiable

The leaderboard hiring managers can actually check

Certs are a red flag without proof. A public ranking of demonstrated red-team solves is the proof. Climb it in real time as you earn marks.

#OperativeGuild Marks
Layer 2 · The flagship credential

The Guild Seal

Guild Marks are your hands-on proof: real work, earned as you go, but taken on your own. The Guild Seal is a different thing entirely. It's a proctored, practical exam with a deliberately low pass rate, one for each track, that you have to sit and pass. The marks get you in the door. The Seal is the one an employer treats as verified.

Proctored
Practical exam
<15%
Target pass rate
Per track
One Seal each
We keep the two apart deliberately. The marks are free, unproctored, and made to be shared, which is what keeps them spreading. The Seal is proctored and hard to get, which is what keeps it respected. Blur them together and both lose their value.
Link copied