Browser AI Kit runs model interactions in the user's browser, but application code, prompts, tool implementations, and rendered output still require normal security review.

Security fixes are released for the latest published minor version of each package.

Please do not open a public issue. Email [email protected] with:

• the affected package and version;
• a concise description and impact assessment;
• reproduction steps or a minimal repository;
• any known workarounds.

We aim to acknowledge reports within three business days and will coordinate disclosure after a fix is available. Good-faith research that avoids privacy violations, service disruption, and data destruction is welcome.

• Chrome's built-in AI output is untrusted data. Escape or sanitize it before rendering HTML.
• WebMCP tools can perform application actions. Re-check authentication and authorization inside every execute function, validate inputs, expose the least data possible, and use accurate tool annotations.
• Do not place secrets in prompts, client bundles, tool descriptions, or client-side configuration.
• Browser AI Kit does not transmit prompts to a DeSource Labs service. Chrome controls model installation and execution; review Chrome and your organization's policies for your deployment.