Ownership
Confirm the partner controls or is authorized to test the target.
Security and responsible testing
Responsible testing with explicit boundaries.
Authorize the scope, isolate sensitive access and stop production runs before irreversible actions.
Authorization starts before the first request.
We record who verified the website, which pages are in scope and what our checker and AI agents may do.
Scope
List the website, pages and actions the run may use.
Actions
Separate read, reversible write and irreversible actions before execution.
Stop gate
Define the exact state where a public or production run must stop.
The environment determines the allowed depth.
A result always carries its audit mode, fixtures and authorization context.
Public
Read public pages, follow links and use site search and filters. Stop before login, account creation, cart, checkout, payment or any form that sends data.
Authenticated
Coming later. Use approved test accounts and scoped permissions, never production customer data.
Checkout rehearsal
Coming later. Only on a store whose owner verified it and approved checkout rehearsal. A made-up test shopper, and a payment provider's published test card only while that provider is in test mode. Code stops the run before the order is placed.
Production stop gates are part of the test contract.
Our checker and AI agents stop before these points, and code enforces it, not the AI model. Checkout rehearsal, coming later, may fill a checkout with test data but still stops before the order is placed.
- Payment or purchase confirmation
- Booking or appointment confirmation
- Lead, application or account submission
- Deletion, cancellation or destructive update
- External message or notification
- Change to production customer data
Controls protect the run and its evidence.
Security state is reported honestly, including limitations and controls that are still being operationalized.
Isolation and egress
Limit reachable targets and external destinations to the approved scope.
Secrets
Keep credentials server-side, scoped and out of screenshots, logs and prompts.
Data minimization
Collect only evidence needed to reproduce and verify the finding.
Redaction
Mask sensitive fields such as passwords, payment details and personal data in step screenshots.
Retention
Step screenshots expire after 30 days on a paid plan and raw page evidence is deleted after 90 days. A workspace is deleted 12 months after its plan ends.
Observability
Record which AI agent and model ran, each step it took and what it was stopped from doing.
Evidence should explain the result without leaking the user.
The collection level is selected before the run and reduced when a lower level can prove the same assertion.
Capture
Keep only step screenshots, step logs and page responses.
Minimize
Remove unrelated content and avoid broad account or page exports.
Redact
Transform sensitive fields before an artifact enters a report or shared workspace.
Retain
Keep the artifact only for the agreed operational and review period.
Report security issues with minimal sensitive detail.
Use the published security contact, describe the affected asset and wait for a secure exchange path before sending secrets or personal data.
- Security contact: security@aiscovery.io
- Policy and canonical contact are published in /.well-known/security.txt
- Do not include credentials, tokens or customer data in the first message
- Testing still requires explicit authorization
No unverified compliance claims.
Aiscovery doesn't claim SOC 2 or ISO 27001 certification, and we don't describe hosting safeguards until our hosting provider is chosen. Our sub-processor list names each provider with its status: in use today, planned or coming later.
The rules in full
Security questions
Can Aiscovery test a site we do not own?
Only with the owner's permission. Our Acceptable Use and Responsible Testing Policy has the rules, and site owners can opt out at any time.
Do public audits make purchases or bookings?
No. They stop before irreversible confirmation or payment.
Can sensitive screenshots appear in a report?
Step screenshots are masked before they are stored or shown to a model: form fields, embedded frames and text that looks like an email, phone, card, account or order number are hidden. Coming later: GPT Work uses OpenAI's own browser and screenshots, which we can't mask.
Define the boundary before testing the journey.
Apply with an owned workflow, approved environment and the constraints your security team needs reviewed.