Review an AI-generated canvas as both a visible artifact and a source project. A clean-looking preview does not tell you everything the files will do when scripts run.
A canvas might show a project timeline, a dashboard, or a visual briefing. It can also reference remote images, include links, or contain code that sends information to another destination. Security review needs to consider those behaviors alongside the design.
Begin with the static view
The current Skybits Guide describes display-only canvas previews by default. HTML and CSS render while canvas-authored scripts, forms, dialogs, and nested frames remain inactive until Allow JavaScript is enabled.
It also says that editing the canvas turns JavaScript off again. That creates an opportunity to inspect changed source before running the revised code. Confirm this behavior in the build you use.
A useful canvas should communicate its main information without requiring you to enable scripts immediately. Ask the agent to make the first view readable and explain what interactivity adds.
Read the destinations
Inspect external links and resources. Check whether a remote domain belongs to an intended source, whether an image URL is necessary, and whether a script calls an endpoint appropriate for the task.
If you cannot assess unfamiliar code, ask someone with the relevant expertise to review it or keep the artifact static. A second explanation from an agent can help identify questions; it is not proof that execution is safe.
Do not place passwords, API keys, or private tokens in canvas source. Anyone authorized to inspect or export the source may receive them, and browser requests can expose them beyond the intended audience.
Keep authenticated services separate
The Guide says public web resources can load in the canvas preview, while Skybits account-only resources and platform-authenticated URLs do not load inside its sandboxed frame.
That means the sandbox has a boundary, but it does not mean the artifact has no network behavior. A static image or external font can still involve a public service. Consider whether the resource is appropriate for the audience and sensitivity of the canvas.
If the design needs private data, use an approved integration pattern and confirm its access model with the team. Do not improvise by embedding a credential in an HTML file.
Check the meaning of the output
Review displayed figures, labels, dates, and footnotes against the source. A dashboard can make a wrong calculation appear convincing, and a chart can imply certainty the underlying data does not support.
Compare the changed source against the prior baseline and inspect the rendered result. Pay attention to removed qualifications and newly introduced claims, not just visual differences.
For example, if the agent converts a forecast into a canvas, make sure “estimated revenue” has not become “revenue,” and that the time period and currency remain visible.
Enable only the behavior you understand
Before enabling JavaScript, ask:
- Which interaction requires it?
- Which scripts and external resources are involved?
- Does any action collect, submit, or transmit information?
- Is the behavior appropriate for everyone who will receive the canvas?
- Have the changed source and visible output both been reviewed?
Canvas source changes are saved before review, so avoid presenting an unreviewed artifact as approved team output. History and baseline comparison help inspect the changes; they do not replace the review itself.
Next step: Keep the preview static, inspect the source and external references, and then enable only the behavior you need. See Skybits Help.