The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes the implementation of interactive playgrounds that ingest and process untrusted code or data from users, creating a surface for injection attacks.
Ingestion points: Detailed in references/detailed-guide.md, playgrounds ingest user input (code, text, API parameters) for immediate execution or processing.
Boundary markers: The guide recommends clearly labeling results as simulated or live, using synthetic input fixtures, and avoiding the inclusion of secrets in shared URLs or analytics.
Capability inventory: The architecture includes executing user code via client-side bundlers (esbuild-wasm), browser environments (WebContainers), or server-side sandboxes (Firecracker, gVisor).
Sanitization: Instructions specify rendering outputs using textContent instead of unsanitized HTML, utilizing content scanners for output safety, and enforcing strict network egress limits.
[DYNAMIC_EXECUTION]: The core functionality described involves the generation and execution of code at runtime within a sandbox environment.
The guide discusses the use of eval(), client-side bundling of user scripts, and server-side containerized execution of arbitrary user-submitted code.
It references several well-known sandboxing and virtualization technologies (Firecracker, gVisor, StackBlitz) to manage the risks associated with this dynamic execution, noting that ordinary containers are insufficient as code boundaries.

developer-sandbox