Architecture facts, not badges.
We hold no third-party security certifications today, and we are not going to imply otherwise. Here is what the system actually does instead.
Governance lives outside the model.
The controls that matter are code that runs before and after the model, not sentences written into a prompt. A prompt can be talked out of a rule. A gate in the request path cannot.
Before the model is called
A deterministic scope screen and a prompt-injection screen run first. Off-topic and adversarial requests are refused without the model ever seeing them.
Only registered material
Retrieval runs against the pages crawled from your nominated domain and recorded in the provenance registry, with the URL, fetch date and hash of each document.
The gate
An answer without registered sources attached does not ship. The widget declines and gives the reason. This is an invariant in the platform, enforced on every request.
Keys and isolation
- Origin-bound keys
- A widget key is issued against the domain you nominate and is bound to that origin. The key is visible in your page source by design, because on its own it does not serve answers anywhere else.
- Per-tenant environments
- Every subscription runs a preprod and a prod environment, each with its own corpus, its own keys and its own bindings. Preprod is where you test a crawl before you accept it; prod is what your visitors see.
- Keys at rest
- Keys are stored hashed, not in plain text, and can be rotated from the portal.
- The edge
- Customer traffic reaches the platform through a hardened edge behind Cloudflare, which carries DNS and the network path in front of us.
No data leaves the platform
Questions are processed on AnswerHatch-operated infrastructure, on machines we run. No third-party model API is in the path of an answer, so a question is never handed to an outside provider in order to be answered. Visitor questions are retained for 30 days and then deleted. The full detail is on the privacy page.
Models are not fine-tuned on a customer's content or on visitors' questions.
The provenance registry
Every crawled document is recorded with the URL it came from, the date it was fetched and its hash. The citations a visitor sees in the widget are a view of that registry. The registry itself is an artifact a compliance officer can ask for, which is a different offer from citations as an interface decoration.
A crawl that fails verification produces a diagnostic. It never becomes a corpus quietly marked live.
Change control and testing
A governance test suite gates every release. The refusal behavior is part of the acceptance criteria for a release, not a property we hope holds.
What we do not claim
We hold no third-party security certification and no independent audit report today. Every competitor page in this category carries the same three badges, and buyers have learned to discount them. When we hold one, this page will name the standard, the auditor and the report date, and you will be able to ask for the report.
Until then, the falsifiable claim on offer is the one on the front page: ask the widget something outside its sources and watch it decline. That is the security property, demonstrated rather than asserted.
Responsible disclosure
If you find a vulnerability, send it to [email protected] with enough detail to reproduce it. We will acknowledge the report and tell you what we are doing about it. We will not pursue good-faith research that stays within these lines: no access to other customers' data, no degradation of service for anyone else, no social engineering of our staff or customers, and no public disclosure until a fix is out.
Buying for a government body, a university or a regulated organization? Send the questionnaire to [email protected] and we will answer it as written, including the questions where the answer is no.