Codex Cryptum 2026

A prompt-injection CTF built with Next.js, PostgreSQL, Drizzle, Better Auth, and the Vercel AI SDK. Uses pnpm workspaces.

Run

pnpm install
cp .env.example .env   # skip if .env already exists
# Fill in the settings below.
pnpm db:migrate
pnpm db:seed
pnpm dev

Open http://localhost:3000. Sign in with Google, then create your team or join using a teammate’s access code. No demo teams are seeded.

Use Node.js 26 and pnpm 12.11.2. The app uses Next.js 16.4.1, React 19.3.1, and TypeScript 7. Dependencies are recorded in pnpm-lock.yaml.

Settings

Keep these in the root .env. Never commit it.

Setting	Purpose
DATABASE_URL	PostgreSQL connection string
BETTER_AUTH_URL	App URL; locally http://localhost:3000
BETTER_AUTH_SECRET	Random secret, at least 32 characters
GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET	Google OAuth web client
LLM_API_KEY	Model provider key
LLM_BASE_URL	OpenAI-compatible base URL, e.g. https://api.openai.com/v1
LLM_MODEL	gpt-6-luna (the event model), or qwen3:8b for the self-hosted fallback (see docs/ENDPOINT.md)
FLAG_SECRET	Independent random secret, at least 32 bytes
ORGANIZER_TOKEN	Independent secret for resetting the sandbox

Register this redirect URI in your Google OAuth client:

http://localhost:3000/api/auth/callback/google

For deployment, use your HTTPS domain in both BETTER_AUTH_URL and the Google redirect URI. Missing credentials produce an error, not a demo fallback.

Cloudflare deployment

Cloudflare Workers Builds deploys main to https://codex.acmvit.in. The connected GitHub repository is ACM-VIT/codex-cryptum-2026; the build command is pnpm build:cloudflare and the deploy command is pnpm --filter cryptum-cloudflare deploy. Builds use Node.js 26 and pnpm 12.11.2.

infra/cloudflare/cloudflare.config.ts declares the Worker, custom domain and Hyperdrive binding. Hyperdrive uses the existing Neon database with query caching disabled so scores and challenge state stay current. Database clients and Better Auth adapters are scoped to each Worker request. Local commands continue using DATABASE_URL from .env.

The OpenNext build uses webpack and a pinned adapter patch that inlines Next.js 16.4's preview-props.json. The patch prevents a runtime manifest error after an otherwise successful build. apps/web/wrangler.jsonc supplies adapter build settings; deployment uses the cf CLI in the separate infrastructure workspace.

To deploy locally, run pnpm deploy after authenticating cf. Application secrets are already stored as encrypted Worker secrets. To replace them from the ignored root .env, run:

pnpm build:cloudflare
pnpm --filter cryptum-cloudflare exec cf deploy --prebuilt --secrets-file ../../.env

Google OAuth must allow https://codex.acmvit.in/api/auth/callback/google. Never commit .env or paste its values into build commands. The Python classifier remains a separate authenticated HTTPS service configured through the Worker secrets.

Model hosting

The event uses OpenAI: set LLM_API_KEY, keep LLM_BASE_URL=https://api.openai.com/v1 and LLM_MODEL=gpt-6-luna. The self-hosted Qwen setup below is a fallback only.

• Backend team: docs/ENDPOINT.md, how to call the model.
• Event day: docs/RUNBOOK.md, setup, fixes and backup.
• Levels 3 and 9: docs/LEVELS.md, how they work and test results.

Layout

apps/web             UI, API routes, authentication, challenge execution
packages/challenges  Public catalog and server-only defense policies
packages/contracts   Shared TypeScript types and Zod schemas
packages/db          Drizzle tables, migrations, development seed

See docs/challenges.md to understand or add a challenge. docs/integration.md lists the original contributions. Component and font credits are in apps/web/THIRD_PARTY.md.

Checks

pnpm typecheck
pnpm lint
pnpm test
pnpm build
pnpm test:db   # optional; use a disposable development database

Database tests temporarily edit sandbox state and restore it. CI runs the first four checks.

Database

• pnpm db:generate: generate SQL after a schema change; review it before applying.
• pnpm db:migrate: apply committed migrations.
• pnpm db:seed: synchronize challenge metadata without creating teams. Existing points, availability, and progress are preserved.

The provisioned Neon database is project dry-band-16656722, branch development, database cryptum, region aws-ap-southeast-1, PostgreSQL 17. Its connection string is only in the ignored local .env. Migrations and seed have been applied.

Safety: the leaderboard sandbox is intentionally vulnerable. Sandbox edits cannot directly change official scores. Challenge filters are puzzle mechanics, not protections for real secrets.

Level 7/8 Python guardrail component

Akhand's framework-independent Python package lives in security/: independent CPU input classification for Levels 7 and 8, Level 8 output classification, team-specific HMAC flags, trusted system prompts and fail-closed policy enforcement. Model weights are pretrained and unchanged. The package targets Python 3.11/3.12.

See security/README.md for the package layout and interfaces, security/DEPLOYMENT.md for model preparation/lifecycle and integration contracts, and security/PORTABILITY_REPORT.md for platform evidence and limitations.

uv sync --locked --python 3.12 --extra ml --extra dev
uv run --no-sync python -m pytest -q -m "not integration"
uv run --no-sync ruff check security
uv build --wheel

The Next.js application and Python guardrails use separate configuration schemas: root .env belongs to the application; optional security/.env contains only the Python keys from security/.env.example. Both local files must remain untracked. Levels 7–8 now use the private pretrained-classifier bridge through a server-only Next.js client. Start it with uv run --no-sync python -m security.classifier_service --env-file security/.env in a separate terminal before admitting participants. Set the same CRYPTUM_CLASSIFIER_TOKEN in root .env and security/.env, and set CRYPTUM_CLASSIFIER_URL in root .env to its private endpoint. Keep the service on loopback or behind authenticated private HTTPS. See security/DEPLOYMENT.md. The existing Next.js provider settings do not automatically change the native Python inference adapter or HMAC derivation. The Python pipeline now supports both Ollama and the official OpenAI SDK; see provider setup. The Python work does not implement authentication, scoring, or any other challenge levels.

Leaderboards refresh every three seconds and immediately on reconnect or tab focus. Official scores come from recorded solves. Level 10 is hidden from the catalog and found through occasional assistant hints; it uses isolated copies of registered teams, initialized from their real scores; practice edits persist until reset. At least two registered teams are required.