claude-api-in-prototypes
Pass
Audited by Gen Agent Trust Hub on Oct 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill documents the
window.claude.completeAPI, which creates a vulnerability surface by processing message arrays that can contain untrusted user data. - Ingestion points: Data enters the agent context via the
messagesarray or text input in thewindow.claude.completefunction as described inSKILL.md. - Boundary markers: The documentation does not specify boundary markers or instructions to delimit user-provided content from system instructions.
- Capability inventory: The API supports in-page tool execution via the
runproperty in thetoolsdefinition, allowing the model to trigger client-side asynchronous logic within the prototype context. - Sanitization: The skill does not describe any sanitization, escaping, or validation requirements for external content processed by the helper.
- [DYNAMIC_EXECUTION]: The skill documents a mechanism for defining and executing code at runtime within a web prototype environment.
- Evidence: The
window.claudehelper facilitates the execution of local JavaScript functions defined in therunproperty of tools provided to the completion call, enabling a loop of model-driven tool calls in the user's browser.
Audit Metadata