OASIS

Offensive AI Security Intelligence Standard — Open-source AI security benchmarking.

Benchmark how AI models perform offensive security tasks — vulnerability discovery, exploitation, privilege escalation, and more. Full analysis with MITRE ATT&CK mapping, behavioral scoring, and detailed reports. Everything runs locally with your own API keys. No account required, no data leaves your machine.

Why OASIS?

AI models are increasingly capable at offensive security. We need reproducible, transparent visibility into how they perform — not behind closed doors, but in the open where the security community can verify, contribute, and improve.

OASIS provides:

• Standardized challenges in Docker containers (CTF-style, isolated, reproducible)
• Multi-provider benchmarking across Claude, GPT, Grok, Gemini, Ollama, and custom endpoints
• Automated analysis with MITRE ATT&CK mapping, OWASP classification, and behavioral scoring
• The KSM scoring model that combines methodology quality with success rate

Quick Start

Prerequisites

• Node.js >= 18
• Docker Desktop (running)
• An API key from any supported provider (or Ollama for local models)

Install & Run

npm install -g @kryptsec/oasis

# Launch interactive mode — walks you through everything
oasis

Or use the CLI directly:

# 1. Set your API key
oasis config set api-key anthropic sk-ant-xxx

# 2. Clone challenges
git clone https://github.com/kryptsec/oasis-challenges.git challenges

# 3. Start a challenge environment
cd challenges/gatekeeper && docker compose up -d && cd ../..

# 4. Run a benchmark
oasis run -c gatekeeper -m claude-sonnet-4-5-20250929 -p anthropic

# 5. View results
oasis results list
oasis report <run-id> --format md

# 6. Share results
oasis report <run-id> -f share --clipboard    # Copy markdown share card
oasis report <run-id> -f html -o report.html  # Standalone HTML report

How It Works

┌─────────────┐     ┌──────────────┐     ┌──────────────┐     ┌────────────┐
│  Challenge   │────>│   AI Agent    │────>│   Analyzer   │────>│   Report   │
│  (Docker)    │     │  (LLM + Kali) │     │  (LLM Judge) │     │  (KSM/ATT&CK)│
└─────────────┘     └──────────────┘     └──────────────┘     └────────────┘

1. Challenge — A Docker environment with a vulnerable target and a Kali attack container
2. Agent — The AI model executes commands in Kali, attempting to find and exploit vulnerabilities
3. Analyzer — A separate LLM evaluates the transcript: technique quality, efficiency, adaptability
4. Report — Scored results with MITRE ATT&CK mappings, OWASP classifications, and KSM rating

Available Challenges

Challenges live in a separate repo and are community-contributed:

You can also create your own challenges.

Scoring (KSM)

The Kryptsec Scoring Model combines methodology quality, success rate, and token efficiency:

Efficacy gating:

The result is then multiplied by the token efficiency factor. Models that burn excessive tokens per step get penalized — up to 30% at extreme inefficiency. Below the 1500 tokens/step baseline, no penalty applies.

Each run also gets a detailed rubric breakdown: objective scoring (flag capture, time/efficiency bonuses), milestone tracking, qualitative assessment, and penalties.

See KSM-SCORING.md for the full specification.

Supported Providers

Model lists are fetched live from provider APIs when an API key is configured. Fallback example lists are shown when no key is available.

Aliases: claude → anthropic, grok → xai, gemini → google

Commands

Run any command with --help for full options.

Analysis

After each benchmark, OASIS uses an LLM (Claude Sonnet by default) to produce:

• MITRE ATT&CK Mapping — Each step classified to specific techniques and sub-techniques
• OWASP Top 10 Classification — Vulnerabilities mapped to OWASP 2021 categories
• Attack Narrative — Executive summary and detailed walkthrough
• Behavioral Analysis — Approach classified as methodical, aggressive, exploratory, or targeted
• Rubric Scoring — Objective metrics, milestone tracking, qualitative assessment, penalties

Analysis uses your Anthropic API key by default. To use a different provider for benchmarking while keeping Anthropic for analysis:

oasis config set api-key anthropic sk-ant-xxx     # For analysis
oasis run -c gatekeeper -m gpt-4o -p openai       # Benchmark with OpenAI

Configuration

Config is stored in ~/.config/oasis/ (XDG-compliant):

• config.json — Settings (default model, provider, paths)
• credentials.json — API keys (local only, restricted permissions, never transmitted)

Environment Variables

Creating Challenges

Challenges are Docker-based CTF environments. Each challenge needs:

• challenge.json — Metadata, scoring rubric, flag, and target info
• docker-compose.yml — Target service + Kali attack container

cp -r challenges/_template challenges/my-challenge
# Edit challenge.json and docker-compose.yml
oasis validate challenges/my-challenge

See the full Challenge Specification and existing challenges for examples.

Development

git clone https://github.com/kryptsec/oasis.git
cd oasis
npm install
npm run build

# Run locally
node dist/index.js --help

# Dev mode (tsx, no build step)
npm run dev -- run -c gatekeeper -m claude-sonnet-4-5-20250929

# Tests
npm test

Contributing

Contributions are welcome! Whether it's new challenges, provider support, bug fixes, or documentation:

1. Fork the repo
2. Create a feature branch
3. Make your changes
4. Run npm test to verify
5. Open a PR

For challenge contributions, submit to oasis-challenges.

License

MIT — Kryptsec