Please do not open a public issue for security vulnerabilities. Report them privately via GitHub Private Vulnerability Reporting (Security → Report a vulnerability).

You can expect:

• An acknowledgment within 48 hours.
• A timeline for the fix and disclosure, typically coordinated within 90 days.

PromptForge is a client-side only application. No data ever leaves your browser except when you explicitly generate prompts against an AI provider you configure.

• API keys are stored locally in your browser's IndexedDB, encrypted with AES-GCM 256 (Web Crypto). The encryption master key is a WebCrypto CryptoKey with extractable: false, also persisted in IndexedDB — key material cannot be exported, so a copy of your browser profile does not directly reveal the key. This is encrypted-at-rest, not zero-leak.
• This is not a secure vault. Any script running in the same origin (e.g., an XSS vulnerability or a compromised extension) can read the key object from IndexedDB and decrypt your data in-page. extractable: false only blocks offline extraction of the key bytes, not in-page decryption. A strict Content Security Policy is the remaining first line of defense against XSS.
• CSP posture: script-src is strict ('self' — no inline scripts, no external script hosts except Vercel Analytics). connect-src allows any HTTPS origin by design: PromptForge is a bring-your-own-endpoint app (OpenAI, Gemini, OpenRouter, or any custom HTTPS endpoint/tunnel you configure), and a static host allowlist would break that feature. Plain http:// is blocked (mixed content) except localhost in development. Note that this means a successful script injection could exfiltrate data to any HTTPS host — the strict script-src is the control that matters.
• IndexedDB is not synced to the cloud by browsers (unlike localStorage), so cloud-sync leakage does not apply. However, a device thief who can run JavaScript in your profile could still decrypt.
• The only true zero-leak option is the "Don't remember API key between sessions" setting on the Settings page: with it enabled, the API key is never written to disk — it lives only in memory for the current session. Endpoint and model are still remembered.
• Prefer short-lived keys, and treat your configured providers as trusted endpoints. API keys are sent only to the provider you configure (OpenAI, Gemini, OpenRouter, or a custom endpoint) via HTTPS.

Only the latest release is actively supported. We do not backport security fixes to older versions.

• Never commit secrets, tokens, or .env* files (they are gitignored — keep it that way).
• Never log API keys, tokens, or prompt content to the console, even in DEV mode.
• All outbound requests must use HTTPS; new endpoints must be validated against the allowed-protocols check in src/lib/.
• User input (prompts, folder names, settings) must be rendered as text, never as HTML.
• Schema/version changes to IndexedDB must ship with a forward-compatible Dexie migration (see src/services/storage/).