CogniScale Trust Centre

Network Sunday Global Limited (company number 07832813, trading as CogniScale)

Registered office: Queensbury House, 106 Queens Road, Brighton, BN1 3XF.

This Trust Centre sets out how CogniScale protects your data and your credentials. It is written for the people who review supplier security, from CISOs to IT leads. For a technical walkthrough, an architecture diagram, our Data Processing Agreement or answers to your security questionnaire, email [email protected].

Clients and prospects can also request access to our in-depth security site, security.cogniscale.com, which covers each of these areas in more detail.

At a glance

• Hosted in the UK and EU. The application database is in London, and our AI routing and supporting services run in the EU (Amsterdam). AI processing by Anthropic and OpenAI, and the knowledge files the Helper syncs to our code-hosting provider, are covered by the transfer mechanisms on our sub-processor page.
• No training on your content. CogniScale never trains on your content, and our AI providers' business terms exclude training on content we send through our business accounts. If you connect your own AI subscription, its own settings apply (section 6).
• Encrypted in transit and at rest. Every connection to our hosted services uses HTTPS. Connected-account tokens, your organisation's AI keys and our off-site backups are encrypted, with the keys held separately from the data.
• Tightly limited staff access. Access to customer environments is restricted to a small number of authorised CogniScale administrators, each using two-step sign-in.
• Tested. We completed an internal security scan of the application, its AI router and the CogniScale Helper on 1 October 2026. The significant findings were fixed within a day, and every remaining finding is tracked to completion. We are preparing for the independent assessment Google requires for its sign-in (CASA) and expect to begin it in October 2026.
• Clear on regulation. Our EU AI Act position is set out in section 15.

1. Where your data lives

AI processing in the United States is covered by the transfer mechanisms named on our sub-processor list at https://cogniscale.com/sub-processors.

2. The Personal Brain

A user who installs CogniScale Helper can store session memories, notes, drafts and personal knowledge on their own device.

Content selected for an AI task may be processed by that task's AI service, so a file stored locally can still leave the device when it is used. The user backs the folder up to storage they personally control, such as their own iCloud or personal Dropbox, or, on macOS, Time Machine. It should not be stored on a company-administered Google Workspace Drive, where Workspace administrators could read it.

Session backups are also stored on the user's own device.

Our privacy notice explains that locally stored material may be processed when it is used in an AI task. We will explain material changes to customers before they take effect, under the applicable agreement.

3. The CogniScale Helper on each user's machine

The CogniScale Helper runs on macOS and Windows. Both installers are signed: the macOS version is signed and notarised through Apple, and the Windows version is code-signed in the name of Network Sunday Global Limited. The Helper listens only on the user's own computer, never on a public network address.

Built in

• Running the AI runtime with the user's own seat or API credentials (Claude Code)
• Reading and writing Personal Brain files and authorised organisation workspace files on the user's machine
• Syncing authorised organisation knowledge with a private GitHub repository through the user's GitHub account
• Reading data from third-party services only when the user has connected them, such as Gmail, Google Drive, Google Calendar, Slack and Microsoft 365 (Outlook, Teams, OneDrive and SharePoint)

What the local runtime can reach

The Helper runs AI sessions with the access of the person using it: their own work files and the tools they have connected, and nothing beyond what their own computer account allows. Review these permissions during setup.

Some sessions send task content, and in router-based sessions connection information, through CogniScale's routing service. Other sessions call the AI provider directly.

The uninstaller removes the Helper. It asks whether to remove runtime data, and it leaves personal content and per-seat configuration for the user to manage.

Customer security teams can review the Helper's design and build evidence with us under NDA.

4. How your credentials are protected

Your organisation's own AI keys

When your organisation brings its own commercial API key, such as an Anthropic Console key or an OpenAI Platform key:

• Keys are stored encrypted by the routing service. The AI router and its gateway run with our European application hosting provider in Amsterdam (Netherlands), and the main application database is with our managed database provider in London (UK)
• Every read, change and deletion of a stored key is recorded, and the record never contains the key itself
• Organisation admins manage their AI keys in the Control Centre
• The service records model usage per organisation and per model

Your team's own Claude subscriptions

When your team uses its own Claude subscription through CogniScale Helper:

• Direct Claude Code sessions use the local Claude sign-in on the user's own device
• Router-based sessions may send a Claude seat token and task content through CogniScale's routing service. Which path applies depends on your configuration

Billing

CogniScale charges for its software. Supplying and billing AI usage directly is on our roadmap (section 16).

5. How data in transit is protected

Connections to CogniScale's hosted services use HTTPS (TLS). CogniScale Helper also serves a loopback HTTP endpoint on the user's own computer for communication with the local AI runtime; the endpoint is not exposed to the public internet.

6. Agentic defence

CogniScale's product is autonomous agents: AI that can take actions on your behalf. That brings risks that traditional software does not have, and we design for them.

• Bounded by the user's access. The co-CEO and Connected Colleagues act on the user's instructions, within the access the user has connected.
• Recorded. Helper sessions are saved on the user's device, and a summary of each session is recorded in the Control Centre. Model usage is recorded per organisation and per model.
• Checked before it is shared. Selected ingestion paths run sensitivity checks before shared knowledge is written to the organisation's wiki.
• Clear on model training. When CogniScale sends Customer content to an AI provider using CogniScale's own business accounts, Anthropic's and OpenAI's published business terms say they do not use that content to train their models. If a user connects their own AI subscription instead, that subscription's own terms apply, including any training setting the user has chosen. Provider retention and processing region depend on the selected account and route.

The next controls we are adding are listed in section 16.

7. Who at CogniScale can access your data

Access to customer environments is restricted to a small number of authorised CogniScale administrators, each of whom uses two-step sign-in. When the Workspace Brain is active (see section 16), staff access to it will require a separate support grant.

8. Our sub-processors

The full, current sub-processor list, with each one's role, processing location and transfer mechanism, is published at https://cogniscale.com/sub-processors. That page is the authoritative source, so it is not repeated here.

We notify customers of material new sub-processors under the process in the Data Processing Agreement.

9. How we handle a security incident

We publish our incident-response procedure at https://cogniscale.com/security/incident-response. Our Data Processing Agreement at https://cogniscale.com/dpa sets out our breach notification and follow-up commitments.

We follow this regardless of whether the incident originates on our infrastructure, on a sub-processor's infrastructure, or in a connected service.

10. Your control over your data

• Export and deletion. Contact [email protected] to request export or deletion of your organisation's data. Our Data Processing Agreement sets out the active-system and backup deletion periods
• Portability. Your conversation history, workspace knowledge and customisations are yours; we can help you take them with you
• Access controls within your organisation. Access is controlled by organisation membership and feature-specific permissions
• Sub-processor objection. When we propose to add a new sub-processor, you have the right to object under the process set out in the Data Processing Agreement
• AI provider terms. See section 6.
• CogniScale's own training restriction. CogniScale does not use Customer content to train CogniScale's own AI or machine-learning models.

11. Encryption

12. Access controls and identity

The Control Centre uses a managed sign-in service together with our own organisation permissions. Platform administrators sign in with two-step verification.

13. Vulnerability management

We scan our code, its dependencies and the running service. Each finding is fixed or tracked to a recorded decision; the most recent full round was on 1 October 2026.

Report vulnerabilities at [email protected]. Our disclosure policy at https://cogniscale.com/security/disclosure sets out our response commitments in full.

14. Compliance posture and certifications

Google requires a CASA security assessment for applications that use its sign-in with access to Gmail, Drive or Calendar. We are preparing for it now and expect to begin it in October 2026. We will share the assessment letter with customers once it is issued.

We are preparing for UK Cyber Essentials certification for the whole company now and expect to apply in October 2026. We do not currently hold SOC 2 or ISO 27001 certification; if your organisation requires one of these before signing, tell us and we will prioritise it.

15. EU AI Act

CogniScale is operated by Network Sunday Global Limited, a UK company. Where our service is offered or used in the European Union, we apply the EU AI Act.

• Our role. Our assessment is that CogniScale is the provider of the co-CEO and Connected Colleagues: AI assistants built on general-purpose AI models from Anthropic and OpenAI. Organisations that use them at work are deployers. Obligations for the general-purpose models themselves rest with Anthropic and OpenAI.
• Risk classification. Our assessment is that CogniScale is a general-purpose business assistant. It is not designed or offered for any practice the Act prohibits, or for the high-risk uses it lists, such as recruitment, promotion or dismissal decisions, or evaluating individual employees.
• Transparency. The co-CEO and Connected Colleagues are presented as AI throughout the service, so the people using them know they are working with AI.
• AI-generated content. From 2 December 2026, the Act requires AI-generated content to be marked in a machine-readable way. We will meet that requirement by that date.
• AI literacy. We support our customers' teams to use AI well, through onboarding and training.
• Your responsibilities as a deployer. If your organisation publishes AI-generated text to inform the public on matters of public interest, the Act may require you to say so, unless a person has reviewed it and takes editorial responsibility.

We keep this position under review as guidance on the Act develops. Questions: [email protected].

16. What we are adding next

• UK Cyber Essentials certification for the whole company, application expected in October 2026
• Compulsory two-step sign-in for the administrators of every customer organisation
• A fresh two-step check before sensitive administrative actions, such as granting administrator rights
• Per-folder permissions for the Helper, with instant revocation
• Task-scoped tokens for agent actions, and platform-wide audit logging of every access and every credential decryption
• Source tagging of every ingested passage, with instruction-stripping and output validation on every ingestion path
• Per-request sub-processor visibility in the Control Centre's audit log
• The Workspace Brain, a guided enterprise upgrade configured for each organisation and drawing on its members' connected accounts, with no IT set-up
• Marking of AI-generated content under the EU AI Act, by 2 December 2026

17. Standards we are documenting our controls against

We are documenting our security controls against CIS Controls v8, NCSC Cyber Essentials guidance, NIST SP 800-53 and the OWASP application and agentic security standards.

18. What we publish

We update each of these when the position changes materially, and the version and date are shown at the top of the document. We provide the CASA assessment letter, and any later penetration-test summary or SOC 2 or ISO 27001 report, to customers under NDA once available.

19. Who to contact

20. Glossary

• Encryption at rest: stored data is encrypted with a key, so it cannot be read without that key
• Encryption in transit: data is encrypted while it moves across the network, so it cannot be read in transit
• Sub-processor: a third party we use to deliver our service, processing your data on our behalf under their own published terms
• Audit log: a record of who did what, and when, used to detect and investigate incidents
• Least privilege: each person or system has the minimum access needed for its role, which limits the impact of a compromised account
• TLS: the standard protocol for encrypting data in transit. The padlock in a browser address bar shows TLS is in use
• Just-in-time (JIT) token: a credential issued when needed, scoped to a single task and expiring when the task completes
• Indirect prompt injection: an attack that hides instructions inside a document, transcript or email, in the hope that an AI agent will read and follow them

Document control