Security
How Premise AI | Outbounder protects access, credentials, research inputs, and generated sales materials.
Effective July 30, 2026English is the source language for the Premise AI | Outbounder website. Pages are localized into other languages with OpenAI models. Check important information.
Current access
The public proof of concept is available without an account or sign-in. It has no end-user accounts, registration flow, password store, or authenticated-user database.
Human verification
Generation requires server-validated Cloudflare Turnstile. The server checks the token, hostname, action, expiry, and single-use result before any OpenAI request.
API credentials
The OpenAI API key is stored as a server-side Sites secret. It is not intentionally placed in browser code, page HTML, local storage, generated kits, or API responses.
Request protection
- Generation accepts same-origin JSON requests with an 8 KB body limit and bounded field lengths.
- Sales methods, deliverables, and languages are restricted to explicit allowlists.
- URL-like values reject credentials, unsupported schemes, private or reserved literal addresses, and nonstandard ports.
- The application does not directly fetch the submitted brand or product URL. Public retrieval is delegated to the provider’s web-search tool.
Abuse and cost controls
Every generation requires human verification, a unique request identifier, replay protection, job authorization, bounded polling, a fixed per-kit model budget, and an owner pause control. Outbounder does not impose a daily or per-client generation quota. It limits simultaneous provider work and durably queues additional requests instead of discarding them.
Operational records
The database stores request identifiers, provider job identifiers, job status, timestamps, aggregate usage and feedback counters, temporary completed results, and read-only snapshots a user explicitly creates. Each completed result is capped at 320 KB and retained in its generation request record for browser handoff recovery. The provider job identifier is the recovery access value, so anyone with the recovery URL can retrieve the result until the request record is removed. Request records older than seven days are removed during the next accepted generation, so deletion may occur after the seven-day mark. Shared snapshots use random 128-bit identifiers, a separate hashed revocation control, rate limits, and a 30-day expiry. Shared-link rate-limit buckets use a daily SHA-256 hash of the requesting IP; the raw IP is not written to that database.
Generated content
React escapes rendered text by default. Generated links are accepted only when they resolve to public HTTP or HTTPS URLs. Source-backed facts remain separate from hypotheses, and unsupported conditions remain questions or evidence boundaries.
Language processing
English is the source interface and support language. Non-English interface and support copy is translated by Premise AI with OpenAI models. A kit’s selected output language is sent with the research request so generated narrative can use that language; structural data keys remain fixed for validation.
Browser storage
Generated kits and display preferences, including language, are stored in local browser storage and do not synchronize across browsers or devices. A completed result is also retained temporarily server-side for browser handoff recovery. Copy link creates a separate shared snapshot.
Browser safeguards
The application sends a restrictive content-security policy, frame denial, strict transport security, content-type protection, a same-origin opener policy, a restrictive referrer policy, and a limited permissions policy. The recovery identifier remains in the URL fragment until the browser sends it in a same-origin request, keeping it out of the initial page request and referrer. Terminal recovery API responses use Cache-Control: no-store.
Limitations
This proof of concept has not completed an independent penetration test, SOC 2 audit, or formal security certification. Do not submit confidential, regulated, customer-controlled, credential, payment, or special-category personal data.