# QEP! Security and Data Handling

Updated November 18, 2026. This document describes the application controls implemented in this release. It is intended for company administrators reviewing QEP. It is not a certification or a promise that every company’s requirements are met.

## Intended use and permitted information

Use QEP for ordinary project plans, task ownership, dates, and progress. Do not enter confidential or proprietary information, payment card data, passwords, API keys, or other secrets. Apply your organization’s approved-use rules to task titles, descriptions, comments, documents, imports, templates, and AI conversations.

QEP does not automatically classify or block every sensitive value. A usage rule does not prevent accidental entry. An authorized administrator can remove affected content and saved copies. Information already copied to an AI provider, exported file, independent project, or another service requires handling in that system as well.

## Identity and permissions

- Website and native ChatGPT plugin requests use the hosting platform’s authenticated, site-scoped identity. QEP separately checks workspace membership and project permissions on its servers.
- Ordinary members can see only projects assigned to them. Editors can change their assigned projects; viewers can read them. Assigning a name to a task does not grant access.
- The workspace owner can explicitly appoint administrators. Administrators have access to all projects in that workspace and can manage ordinary members, rules, saved data, and connection keys. Only the owner can appoint or change other administrators. The owner cannot be removed by a member-management command.
- New projects are private from ordinary members until an administrator grants access. Workspaces are separated by their owning account’s site-scoped ID.
- Invitations identify the recipient’s QEP account email or QEP user ID, selected projects, and role. They are single use and expire after at most seven days. Old invitations that did not specify a recipient must be replaced. Removal or a permission change invalidates earlier invitations for the account, including unused email invitations. A new invitation can be issued afterward. QEP does not send invitation emails automatically.

Recipients can copy their account email and QEP user ID from **Team & access**. The account email supplied by the identity provider may differ from a person’s usual work address, including when a relay address is used.

## AI connections

The native ChatGPT plugin acts under the authenticated caller. Supported external MCP clients use a workspace-bound connection key supplied in an Authorization header. Raw keys are shown once, stored only as SHA-256 hashes, and never returned by inventory commands. Generated keys have 256 bits of random material and expire after a configurable 1–30 days; each member can have at most five active keys in a workspace.

Connection requests check current membership and project permissions. Administrators can list and permanently revoke workspace keys. Members can revoke their own keys, including after a downgrade to viewer. Revoking a portable key disables that connection. Native ChatGPT uses the signed-in account and has no portable key to revoke. Removing a member, changing their project permissions or administrator role, or accepting an additional invitation permanently revokes that member’s existing workspace keys. Rejoining requires a fresh key; revoked keys do not resume working. A key is a bearer credential: protect it and revoke it if exposed.

Disabling AI project access blocks project operations through both AI transports, blocks new connection keys, and permanently revokes existing keys. The website remains usable. Authorized security administration through the native ChatGPT plugin remains available to recover or change company settings. External key-based connections stay disabled until access is enabled and a new key is authorized.

New AI project setups are saved as drafts and reviewed in QEP before they create a project. Routine edits to project work rely on the chat client for confirmation. Administrative changes proposed through either chat connection require a separate signed-in QEP review.

## Company rules and administration through chat

Administrators can ask a connected AI to read members, connection metadata, protected history, saved-data inventories, or current company rules, and to prepare administration changes. The requester opens the ten-minute review link, checks the exact workspace and parameters, and approves or discards the proposal. The server accepts the original stored arguments once and verifies authority again at execution. A revoked original connection cannot execute a pending change. Invitation links are returned on the approval page. Every command checks the caller’s actual workspace authority. A key cannot target another workspace. No command grants arbitrary access to other companies.

Supported company settings are deliberately small:

| Setting | Built-in default | Effect |
| --- | --- | --- |
| AI project access | Enabled | Controls project operations through the native ChatGPT plugin and key-based AI connections. Security administration through the native plugin remains available. |
| External template sharing | Disabled | Both source and receiving workspaces must enable it. The source administrator must explicitly select a receiving workspace where they have edit access. Turning it off blocks shared-template retrieval. |
| Connection lifetime | 30 days | May be shortened to 1–30 days. Shortening also reduces the expiry of existing keys. |
| Invitation lifetime | 7 days | May be shortened to 1–7 days for new invitations. |

Read current settings and their version before saving changes. Unsupported policy instructions are rejected; telling an AI an arbitrary policy does not cause QEP to enforce it. QEP does not claim to enforce employer device management, company identity-provider MFA, personal AI account policies, or data handling outside QEP.

## Protected history and data integrity

Successful workspace changes, project/task changes made through AI, invitations, membership changes, connection administration, policy changes, draft operations, and template operations create security events. The application records stable actor and workspace IDs, operation, target, timestamp, outcome, and limited change metadata. External AI operations also identify the connection. Browser edits include record IDs and changed field names; the history does not store the full old and new content.

Event creation is in the same database transaction as the associated mutation. Failed writes do not create successful-change events. The application offers no event-editing interface. Database triggers reject event updates and deletion of recent events. Access to history is limited to authorized workspace administrators. Rejected administrator and disabled-AI policy checks also create denial events. These controls do not make database administrators or the hosting provider incapable of modifying infrastructure.

History is retained for a 90-day window, with expired records removed opportunistically during subsequent signed-in use. Cleanup is not a guaranteed scheduled erasure deadline. Events omit raw connection keys, invitation tokens, task/document bodies, and rejected input payloads. Company admins should also avoid sensitive values in record identifiers.

Workspace and template writes use current versions to reject stale changes. Permission and connection state are rechecked in the database mutation to prevent a stale authorization snapshot from committing after revocation.

## Deletion and retained copies

- Project deletion removes its live issues, comments, cycles, and documents from workspace state. Independently saved templates and AI setup drafts remain until explicitly deleted.
- Administrators can delete saved templates, including their shared availability, and pending or applied AI draft copies. Live projects previously created from those copies remain independent.
- The owner can explicitly erase workspace application data through the protected `erase_workspace_data` command. This removes live plans, documents, templates, drafts, invitations, connections, memberships, and company settings, while retaining limited security metadata for its retention window. It does not delete the user’s ChatGPT account.
- Deleting content in QEP does not recall exports or copies in external AI conversations or independently created projects. Provider backup retention and deletion require separate evidence and procedures.
- Expired invitations and connection records are eligible for opportunistic cleanup 30 days after expiry. Request-counter rows are pruned after a day. Revoked keys remain unusable while records await cleanup.

## Request and browser protections

The server enforces bounded request reads, schema and reference validation, per-account request ceilings, and separate quotas for invalid keys and authenticated accounts. Invalid requests do not spend the request allowance of a valid connection. Looking up a presented key still consumes a database read. Protection against traffic that exhausts infrastructure capacity requires provider edge controls. Limits include 60 website workspace saves/minute, 60 membership writes/minute, 30 connection-setting writes/minute, 30 template writes/minute, and 60 security administration requests/minute. Multiple open tabs and AI clients share the relevant account limits. Excess requests receive a retry response. These application limits complement, and do not independently establish, the provider’s edge protections.

Browser write routes reject foreign request origins. Production worker responses add a content-security policy, MIME sniffing protection, a no-referrer policy, HTTPS transport policy, and disabled camera/microphone/geolocation browser permissions. The content-security policy currently permits inline scripts/styles needed by the framework; it is not a nonce-only policy. Effective headers can also be affected by the hosting platform and should be checked in a company review.

## Hosting responsibilities and assurance

QEP runs on the Sites hosting platform with a Cloudflare Worker and D1 storage. The application uses the platform’s identity and managed storage integrations. The publisher’s access and deployment controls are separate from company membership inside QEP.

No QEP-scoped SOC 2 report, ISO/IEC 27001 certificate, guaranteed data-residency commitment, customer-enforced SSO/SCIM integration, or independent penetration-test report is asserted by this document. Encryption at rest, provider key management, storage/processing locations, platform identity-header handling, session controls, backup recovery, incident response, and contractual protections require evidence from the relevant operator or provider. A hosting provider’s certification does not automatically certify the application.

Company security teams should review their data scope, provider agreements, corporate identity requirements, AI account approval, retention requirements, recovery evidence, and applicable laws before approval. Keep regulated data out of QEP unless its specific requirements have been separately assessed and approved.
