If you discover a security vulnerability, please report it responsibly:

Email: hello@herakles.dev

Do not open public issues for security vulnerabilities.

1. Input validation — Content is validated as legal/TOS text before reaching the LLM. Non-legal content is rejected with a descriptive error.
2. System prompt hardening — The Gemini system prompt constrains output format (JSON schema) and explicitly instructs the model to ignore embedded instructions in user content.
3. Output validation — LLM responses are parsed against a strict Zod schema. Any response that doesn't match the expected structure is rejected.

• Nginx layer — IP-based rate limiting at the reverse proxy level
• Redis layer — Atomic Lua scripts enforce per-IP rate limits with sliding windows. No race conditions.

• Daily Gemini API token budget caps prevent runaway costs
• Configurable via DAILY_TOKEN_BUDGET environment variable
• Automatic rejection of requests when budget is exhausted

• Analysis content is hashed (SHA-256) for deduplication — duplicate TOS documents return cached results
• No user accounts or personal data collected
• Analysis results expire after 30 days (configurable)
• Shared links use non-sequential UUIDs

• PostgreSQL with Prisma ORM (parameterized queries, no SQL injection surface)
• Redis connections via ioredis with TLS support
• Docker deployment with non-root container user
• HTTPS enforced via Cloudflare + nginx