Skybits brings people and AI agents into the same workspace. Security starts with clear answers: who can access your work, what an agent can do, where information is processed, and how you review a change.
You can set access for collaborators, limit connected agents, and inspect changes through review tools and history. Before using sensitive information, review both Skybits' data practices and the practices of any AI service you connect.
How Skybits works
Skybits stores collaborative documents, canvases, spreadsheets, and uploaded files. People and agents work on these items under the permissions available to them. Agents can create and edit work, so their access and your review process both matter.
You can use the Skybits cloud agent, connect an external assistant, or use a supported local agent through Agent Bridge. These options have different data paths. An agent running on your computer can still use a remote model provider, and information included in a conversation or tool result can cross the network.
Control access to your work
Choose collaborator roles. Owners, editors, commenters, and viewers have different capabilities. Give each person the role they need for the task. Comment availability depends on the item type; spreadsheets are view-only for commenters in the documented first release.
Review every route to access. A person may have direct access, access inherited from a shared folder or tag, or access through a link. Removing one route may leave another in place. Shared folders can also give access to items added later.
Use link sharing deliberately. Where link access is enabled, an item can be available to anyone with its link at the chosen role. Direct invitations are usually easier to manage for sensitive work. Changing link access does not automatically remove a person's stronger direct grant.
Decide which incoming shares to accept. Skybits can ask you to accept a shared item and lets you trust or block its sender. Where acceptance is required, agents cannot read or edit it until you accept. Trusting a sender does not establish that every instruction or link inside their content is safe.
Give agents only the authority they need
Agent requests are checked against credentials and document permissions. Connector API keys also carry tool scopes, so a client that only needs to read does not need permission to create or edit.
In Settings, review Agent permissions and Connections & API keys. The current Guide describes separate controls for agent sharing, Trash, and version restoration. Agent sharing is off by default. Turning it on permits actions within the limits you set; it does not create a human approval step for every action.
Use an expiry for temporary API keys, keep secrets in an appropriate credential store, and revoke connections that are exposed or no longer needed. A connector's tool scope does not by itself mean it has a separately selected allowlist of documents: document access must also be checked.
Review changes in the form that fits the item
| Item | How agent changes are reviewed |
|---|---|
| Rich-text document | Content edits arrive as suggestions that a person can accept or reject. |
| Spreadsheet | Changes update the live workbook and appear with review markers. Human editors inspect the Unreviewed changes panel; agents cannot accept their own spreadsheet review entries. |
| Canvas | Source changes are saved collaboratively. Review the source, static preview, baseline comparison, and history. |
| Uploaded File | A new version becomes available immediately. Compare or download versions, or restore an earlier version as a new one. |
These controls make changes easier to inspect. They do not guarantee that every action waits for approval, that an AI answer is correct, or that a restored item can retract information already shared elsewhere.
Treat interactive canvases as code
Canvas previews are display-only by default. The current Guide describes an Allow JavaScript control and says that changing the canvas turns scripts off again. Inspect changed source before enabling them.
Public web resources can load in a canvas preview. Review external links, images, fonts, and network destinations as well as visible text. A sandboxed preview does not establish that a canvas makes no external requests.
Protect stored information and connections
The published privacy policy describes encryption in transit and at rest, private network placement for data stores, verified sign-in, hashed connector keys, and restricted production access. These protections address different risks: interception, exposure of stored data, unauthorized connections, and staff access.
Encryption at rest and in transit is not a claim of end-to-end encryption or zero-knowledge storage. Skybits and the services involved in your requested workflow may process authorized content. For requirements involving exact algorithms, key ownership, internal network paths, or independent assurance, ask the team for current evidence.
Understand AI data handling
Using AI can involve document excerpts, prompts, instructions, comments, generated output, and other task context. Hosted conversations and external assistants should be evaluated separately.
The published Skybits privacy policy explains hosted-agent processing and operational data use. If you connect your own assistant, review that provider's terms and your account settings too. A model-training restriction and a retention restriction answer different questions.
The policy says Skybits does not use document content, prompts, or generated output to train Skybits AI models or third-party foundation models without your express permission. An external assistant you choose is subject to its own provider terms. This training commitment does not mean conversations are never stored or requests have zero retention.
For current providers, retention, telemetry, international transfers, and privacy requests, see the Privacy Policy. If your workflow needs a particular processing location or contractual condition, discuss it with the team before uploading sensitive material.
Frequently Asked Questions
Can an agent access every document in Skybits?
Agent access depends on its authenticated identity, document permissions, and the applicable action controls. An agent is not granted access to other users' private documents just by connecting. Review all accessible material, including folder inheritance, before assuming the agent is limited to the document currently open.
Can I connect an assistant only for reading?
The documented API-key setup offers read access separately from create/edit access. Choose the minimum scopes needed. OAuth-connected assistants use the authorizing user's document authority; do not assume every connection type exposes identical scope controls.
Are all AI changes held for my approval?
Rich-text content edits arrive as suggestions. Spreadsheet edits update the live workbook; canvas edits and file versions are saved before review. Sharing, organization, and other management actions have their own permission rules. Review the controls relevant to the task and item type.
Does running a local agent keep all information on my computer?
No. The model process and provider credentials can remain local while chat messages, replies, tool transcripts, and task information cross the network through Agent Bridge. The local runtime may also call a remote model service. Local file and command access depends on that runtime's permissions.
Can I see what an agent changed?
History and review surfaces help you inspect changes and distinguish agent activity from human editing. Depending on the item, you can compare or restore a prior revision. These are useful operational records; a regulated workflow may require additional approval and recordkeeping controls.
What happens when I remove access?
Review direct sharing, inherited folder access, link access, and agent connections. The Guide describes live collaboration updates when permissions change. Removing access limits future activity; it does not retract downloads, exports, screenshots, or information already received by another service.
What happens when I delete a document or account?
Trash, active-system deletion, backups, shared contributions, and provider copies have different lifecycles. For current retention rules and the account-deletion process, consult the Privacy Policy. Document export and a full account-data export are also different operations.
Can Skybits prevent prompt injection?
Untrusted content can contain instructions aimed at an agent. Permission checks and restricted actions reduce what an agent can do, while narrow source instructions and human review help you notice problems. No claim of complete prompt-injection prevention is made here.
Does Skybits hold security certifications?
The published privacy policy states that Skybits does not currently hold SOC 2, ISO 27001, or similar certifications. Ask the team about current assurance evidence if your approval process requires it.
Are enterprise identity, private deployment, and formal approval chains available?
Confirm the specific capability and setup with the Skybits team. Document roles and content review do not establish organization-wide administration, enterprise SSO/SAML, a private deployment, or a formal corporate approval chain.
How do I raise a security or privacy question?
For privacy questions, use the contact details in the Privacy Policy. General enquiries can go to info@skywood.ai. If reporting a suspected issue, describe the affected feature and time without including passwords, API keys, or other people's private content. Ask for an appropriate private channel before sharing sensitive evidence.
Start with a bounded workflow
Choose one task, a small set of approved sources, and the minimum access needed. Name a reviewer. Check sharing before the output leaves the workspace.
Explore the Skybits Help or contact the team to discuss your requirements.