How It Works

Not a course. A living ecosystem.

LevelUpCTF isn't pre-authored content. In the REACTOR v2 lane, a Model Council of specialist agents designs, critiques, and revises through SAGE; independent solver, quality, and operator gates decide what earns promotion. You work it in a real Docker sandbox; the platform recalibrates nightly.

LOOP 01GENERATE

REACTOR drafts each challenge.

The Model Council starts with a specialist Designer: a vulnerable web app, intrusion scenario, malicious binary, or smart contract with a subtle bug. Peers critique the brief through SAGE; Static Analysis catches generator slop before build; Validator and Calibrator iterate toward the target ELO band.

REACTOR v2 treats a draft as a hypothesis, not a release. The exact revision must survive live validation, independent solver and quality review, then operator sign-off before promotion.

  • Designer + Static Analysis ping-pong until the brief stabilizes
  • Target difficulty set from the gap in your skill vector
  • Flag format enforced: LEVELUP{...}
designer.log
▶ designing · category: web · target ELO 1724
→ premise: SaaS vendor portal, IDOR on /invoices
→ twist: logs delayed 4m behind actions
✓ static_analysis passed · 0 lint failures
✓ validator accepted after 2 revisions
→ estimated par 00:42:00
LOOP 02VALIDATE

Six stages plus two branches. Before you ever see it.

Every draft runs through the real state machine in orchestrator.py. We don't ship a broken sandbox. We don't ship an unsolvable one. And we definitely don't ship one an LLM can write up from public data.

STAGE 01
Init
Category locked, challenge_type inferred, target difficulty set.
STAGE 02
Generate
Designer drafts the brief, narrative, Dockerfile, and flag placement.
STAGE 03
Static Analysis
Deterministic linter rejects missing flags, broken Dockerfiles, generator slop.
STAGE 04
Validate
Image builds, services come up healthy, Exploit agent reaches the flag end-to-end.
STAGE 05
Calibrate
Par time measured, difficulty scored against target band.
STAGE 06
Deploy
Hardened container image ships to the library with stream labels.
BRANCH · REPAIR
Patch, don't regenerate
On stage failure, the Repair agent pulls institutional memory and fixes the specific stage — up to two attempts.
BRANCH · DISCARD
Throw it away
If repair runs out or calibration gap exceeds 1.0, the draft is discarded. Lessons recorded for the next generation.
LOOP 03PLAY

You get a real box.

Not a multiple-choice quiz. Not a text-adventure sim. A live Docker container with a full analyst toolkit, a terminal, and a flag that requires real tradecraft to capture.

  • Kali-style terminal + category-specific tools: sqlmap, pwntools, Volatility, Foundry
  • Hints cost coins · streaks protect your XP · Eagles double your loot
  • Clock runs against the AI solver's par — beat it and you pocket the Eagle
sandbox.run
$ whoami
analyst
$ cat /etc/motd
LevelUpCTF · stream: PURPLE · par 00:42:00
ticket: IR-2471 · user reports portal weirdness
$ tail -f /var/log/nginx/access.log
10.77.3.14 - GET /invoices/104892 HTTP/1.1 200
10.77.3.14 - GET /invoices/104893 HTTP/1.1 200
10.77.3.14 - GET /invoices/104894 HTTP/1.1 200
LOOP 04LEARN

Every solve reshapes your skill vector.

Win, lose, or time out — your 15-axis challenge-type ELO updates. A matchmaker queues the next challenge in your growth zone: hard enough to stretch, not so hard you bounce off.

  • Per-category ELO · σ drops as the system learns you
  • Stuck? Targeted learning surfaces the exact resource, then brings you back
  • Rank up · earn a badge · extend your streak · bank coins
skill-vector · last 30 days
Web 1821 ▲ +47
Crypto 1612 ▼ -12
Pwn 1543 ▲ +8
DFIR 1794 ▲ +61
Forensics 1702 ▲ +14
Malware 1189 ▼ -4 ← targeted
Reversing 1334 ▲ +7
...

next match: Malware · medium · +34 expected ELO
NIGHTLYEVOLVE

Five loops at 00:01 UTC.

The evolution worker runs the nightly cron. Challenges that everyone solves too fast get mutated. Prompts that produce boring briefs get retired. Gaps in the catalogue get filled from real solve-rate data.

  • evo.mutate — rewrite easy challenges to be harder-but-different
  • recalibrate — re-score par based on actual solve data
  • prompt.evolve — retire stale archetypes, breed better ones
  • gap.fill — generate new challenges for under-covered skills
  • retire — archive challenges that no longer meet the catalogue standard

You wake up to a smarter platform. Every. Single. Day.

Start Hacking →Pick a Stream
How It Works | LevelUpCTF