Developers have implemented a sandbox environment for an AI manager that writes and executes JavaScript code to interact with marketplace APIs. The code runs locally in the user's browser via a Web Worker rather than on the server, ensuring that sensitive keys and server files remain isolated from the execution process.
The system uses a proxy to handle API requests for Ozon, Wildberries, and Yandex Market. For Wildberries, the proxy verifies methods against official documentation to prevent unauthorized write operations. Ozon and Yandex Market integrations are currently more restrictive, relying on keyword filtering in request paths to distinguish between read and write actions.
Between October 21 and November 1, the system recorded 118 executions across 46 tools in three projects. Ozon accounted for the majority of activity with 72 runs. Analysis shows that nearly half of the tools are used for one-time tasks or API exploration, while others handle recurring analytical functions.
Despite the sandbox, security risks remain. The AI can still exfiltrate data it is legally authorized to access, such as orders or reviews, via network requests. The developers acknowledge that the current blacklist-based protection is incomplete and suggest that a whitelist approach for API methods and network policies is necessary for robust security.
