ai-scanner-garak — Автоматизированный сканер безопасности моделей ИИ

--- name: ai-scanner-garak description: AI model safety scanner built on NVIDIA garak for testing LLMs against 179 security probes across 35 vulnerability families triggers: - scan an AI model for vulnerabilities - test LLM security with garak - run AI safety assessment - set up ai-scanner for penetration testing - configure AI model security scanning - check LLM for OWASP top 10 vulnerabilities - schedule recurring AI security scans - export AI scan results to PDF or SIEM --- # AI Scanner (0din-ai/ai-scanner) > Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection. AI Scanner is an open-source Ruby on Rails web application for AI model security assessments, wrapping [NVIDIA garak](https://github.com/NVIDIA/garak) with a multi-tenant UI, scheduling, PDF reports, and SIEM integration. It runs 179 community probes across 35 vulnerability families aligned with the OWASP LLM Top 10. ## Installation ### Quick Install (Docker) ```bash curl -sL https://raw.githubusercontent.com/0din-ai/ai-scanner/main/scripts/install.sh | bash ``` ### Manual Install ```bash curl -O https://raw.githubusercontent.com/0din-ai/ai-scanner/main/dist/docker-compose.yml curl -O https://raw.githubusercontent.com/0din-ai/ai-scanner/main/.env.example cp .env.example .env ``` Edit `.env` with required values: ```bash # Generate a secure key openssl rand -hex 64 # .env minimum required values SECRET_KEY_BASE=<output_of_above_command> POSTGRES_PASSWORD=<your_secure_db_password> ``` ```bash docker compose up -d ``` Access at `http://localhost` — default credentials: `admin@example.com` / `password`. **Change the default password immediately after first login.** ## Configuration (.env) ```bash # Required SECRET_KEY_BASE=<64-byte-hex> POSTGRES_PASSWORD=<strong-password> # Optional: custom port PORT=8080 # Optional: SIEM integration SPLUNK_HEC_URL=https://splunk.example.com:8088/services/collector SPLUNK_HEC_TOKEN=$SPLUNK_HEC_TOKEN RSYSLOG_HOST=syslog.example.com RSYSLOG_PORT=514 # Optional: email SMTP_HOST=smtp.example.com SMTP_PORT=587 SMTP_USERNAME=$SMTP_USERNAME SMTP_PASSWORD=$SMTP_PASSWORD ``` ## Core Concepts | Concept | Description | |---|---| | **Target** | An AI system to test — API-based LLM or browser-based chat UI | | **Probe** | A single attack test (e.g., prompt injection, data leakage) | | **Scan** | A run of selected probes against a target | | **ASR** | Attack Success Rate — percentage of probes that succeeded | | **Organization** | Tenant boundary; users and scans are scoped per org | ## Setting Up a Target Targets define what you're scanning. Two types: **API-based LLM Target** (e.g., OpenAI-compatible endpoint): ```ruby # In Rails console or via UI — representative model target = Target.create!( name: "Production GPT-4", target_type: "api", api_endpoint: "https://api.openai.com/v1/chat/completions", api_key: ENV["OPENAI_API_KEY"], model_name: "gpt-4", organization: current_organization ) ``` **Browser-based Chat UI Target**: ```ruby target = Target.create!( name: "Internal Chatbot UI", target_type: "browser", url: "https://chatbot.internal.example.com", organization: current_organization ) ``` ## Running a Scan ### Via UI 1. Navigate to **Targets** → select your target 2. Click **New Scan** 3. Select probe families or individual probes 4. Click **Run Scan** ### Via Rails Console ```ruby # On-demand scan with specific probe families scan = Scan.create!( target: target, probe_families: ["prompt_injection", "data_leakage", "insecure_output"], organization: current_organization ) ScanJob.perform_later(scan.id) ``` ### Scheduled Recurring Scan ```ruby # Weekly scan every Monday at 2am scheduled_scan = ScheduledScan.create!( target: target, probe_families: ["prompt_injection", "jailbreak"], cron_expression: "0 2 * * 1", organization: current_organization ) ``` ## Probe Families (35 total, aligned to OWASP LLM Top 10) Key probe families available: ```ruby # List all available probe families Garak::ProbeRegistry.families # => ["prompt_injection", "jailbreak", "data_leakage", "insecure_output", # "supply_chain", "sensitive_info", "excessive_agency", "overreliance", # "model_theft", "malicious_plugins", ...] # Get probes within a family Garak::ProbeRegistry.probes_for("prompt_injection") # => 179 total probes across all families ``` ## Viewing Results ### Attack Success Rate (ASR) ```ruby scan = Scan.find(scan_id) puts scan.asr_score # => 0.23 (23% attack success rate) puts scan.status # => "completed" puts scan.probe_results.count # => 47 # Per-probe breakdown scan.probe_results.each do |result| puts "#{result.probe_name}: #{result.passed? ? 'SAFE' : 'VULNERABLE'}" puts " Attempts: #{result.attempt_count}" puts " ASR: #{result.asr_score}" end ``` ### Trend Tracking ```ruby # Compare ASR across scan runs for a target target.scans.completed.order(:created_at).map do |scan| { date: scan.created_at, asr: scan.asr_score } end ``` ## PDF Report Export ```ruby # Generate PDF report for a scan scan = Scan.find(scan_id) pdf_path = ReportExporter.export_pdf(scan) # Includes: executive summary, per-probe results, per-attempt drill-down ``` Via UI: Navigate to a completed scan → **Export PDF**. ## SIEM Integration ### Splunk ```ruby # config/initializers/siem.rb SiemIntegration.configure do |config| config.provider = :splunk config.splunk_hec_url = ENV["SPLUNK_HEC_URL"] config.splunk_hec_token = ENV["SPLUNK_HEC_TOKEN"] config.forward_on_completion = true end ``` ### Rsyslog ```ruby SiemIntegration.configure do |config| config.provider = :rsyslog config.rsyslog_host = ENV["RSYSLOG_HOST"] config.rsyslog_port = ENV["RSYSLOG_PORT"].to_i end ``` ## Multi-Tenant Organization Management ```ruby # Create a new organization org = Organization.create!(name: "Security Team Alpha") # Invite a user user = User.invite!( email: "analyst@example.com", organization: org, role: "analyst" # roles: "admin", "analyst", "viewer" ) # Data is encrypted at rest per organization org.encryption_key # => managed automatically ``` ## Development Setup ```bash git clone https://github.com/0din-ai/ai-scanner.git cd ai-scanner cp .env.example .env.development # Install dependencies bundle install # Database setup rails db:create db:migrate db:seed # Install garak (Python dependency) pip install garak # Start development server bin/dev ``` ### Running Tests ```bash # Full test suite bundle exec rspec # Specific area bundle exec rspec spec/models/scan_spec.rb bundle exec rspec spec/jobs/scan_job_spec.rb # Lint bundle exec rubocop ``` ## Common Patterns ### Testing a New LLM Before Deployment ```ruby # Comprehensive pre-deployment scan target = Target.create!( name: "New Model v2 - Pre-deploy", target_type: "api", api_endpoint: ENV["NEW_MODEL_ENDPOINT"], api_key: ENV["NEW_MODEL_API_KEY"], model_name: "new-model-v2", organization: current_organization ) # Run all 35 probe families scan = Scan.create!( target: target, probe_families: Garak::ProbeRegistry.families, organization: current_organization ) ScanJob.perform_now(scan.id) if scan.reload.asr_score > 0.15 puts "WARNING: ASR #{scan.asr_score} exceeds threshold. Review before deploying." else puts "PASS: Model meets security threshold." end ``` ### Using the Mock LLM for Testing Scanner Setup The built-in Mock LLM lets you validate your scanner configuration without hitting real APIs: ```ruby target = Target.create!( name: "Mock LLM", target_type: "mock", organization: current_organization ) # Run a quick scan to verify everything works end-to-end ``` ### Webhook on Scan Completion ```ruby # config/initializers/scan_hooks.rb ActiveSupport::Notifications.subscribe("scan.completed") do |_, _, _, _, payload| scan = Scan.find(payload[:scan_id]) if scan.asr_score > 0.20 SlackNotifier.alert( channel: "#security-alerts", message: "High ASR detected: #{scan.asr_score} on #{scan.target.name}" ) end end ``` ## Docker Compose Production Tips ```yaml # docker-compose.override.yml — production additions services: web: environment: RAILS_ENV: production FORCE_SSL: "true" labels: - "traefik.enable=true" - "traefik.http.routers.scanner.rule=Host(`scanner.example.com`)" - "traefik.http.routers.scanner.tls.certresolver=letsencrypt" ``` ```bash # Upgrade docker compose pull docker compose up -d docker compose exec web rails db:migrate ``` ## Troubleshooting | Problem | Solution | |---|---| | Scan stuck in "running" | Check `docker compose logs worker` — garak Python process may have crashed | | `SECRET_KEY_BASE` error on start | Run `openssl rand -hex 64` and set in `.env` | | Can't connect to target API | Verify API key env var is set; check firewall allows outbound from container | | Browser target scan fails | Ensure Playwright/Chrome is available in the worker container | | PDF export blank | Check `wkhtmltopdf` is installed in the web container | | SIEM not receiving events | Verify `SPLUNK_HEC_URL` includes full path `/services/collector` | ```bash # View all service logs docker compose logs -f # Check worker specifically (runs garak) docker compose logs -f worker # Rails console for debugging docker compose exec web rails console # Check garak is working docker compose exec worker python -c "import garak; print(garak.__version__)" ``` ## Key Files (for Contributors) ``` app/ models/ scan.rb # Core scan model, ASR calculation target.rb # Target types and validation probe_result.rb # Per-probe result storage jobs/ scan_job.rb # Async job that invokes garak services/ garak_runner.rb # Ruby wrapper around garak CLI report_exporter.rb siem_integration.rb lib/ garak/ probe_registry.rb # 179 probes, 35 families dist/ docker-compose.yml # Production compose file scripts/ install.sh # One-line installer ``` ## Resources - [Full Documentation](https://0din-ai.github.io/ai-scanner/) - [Quick Start Guide](https://0din-ai.github.io/ai-scanner/getting-started/quick-start) - [NVIDIA garak](https://github.com/NVIDIA/garak) - [OWASP LLM Top 10](https://owasp.org/www-project-top-10-for-large-language-model-applications/) - [CONTRIBUTING.md](https://github.com/0din-ai/ai-scanner/blob/main/CONTRIBUTING.md) - [SECURITY.md](https://github.com/0din-ai/ai-scanner/blob/main/SECURITY.md)