The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents the integration of external web content which introduces a surface for indirect prompt injection or XSS from untrusted websites.
Ingestion points: Data is ingested through URL requests and HTML loading methods in SKILL.md and references/loading-and-observation.md.
Boundary markers: The instructions recommend restricting navigation to trusted domains using WebPage.NavigationDeciding in references/navigation-and-javascript.md.
Capability inventory: The skill provides instructions for JavaScript execution via callJavaScript and local content serving via custom schemes in references/local-content-and-custom-schemes.md.
Sanitization: The documentation explicitly promotes the use of the arguments dictionary for JavaScript calls to prevent injection via string interpolation in references/navigation-and-javascript.md.
[DYNAMIC_EXECUTION]: The skill describes the use of callJavaScript to execute dynamic code within the web view context.
Patterns: Documentation in SKILL.md and references/navigation-and-javascript.md shows how to evaluate JavaScript strings at runtime.
Mitigation: The skill emphasizes using standard API arguments to safely pass data between Swift and JavaScript environments, reducing the risk of malicious code injection.

swiftui-webkit