Gemini Spark Leak: Full-time Back Office Agent May Place Trades Without User Consent

According to QuantAI Beating monitoring, Google is conducting internal testing of a 24/7 AI Agent named "Gemini Spark" on the Gemini web version. Developer @mrfanduuuuu today leaked a screenshot of its onboarding page on X, marked as BETA, with an expected official unveiling during the May 19 I/O keynote.

Spark's core selling point is "the more you use, the more it understands you." It will extract context from connected Google apps, skill modules, chat logs, scheduled tasks, logged-in websites, Personal Intelligence (Google's personal data integration feature launched in January this year), and location information to perform multi-step tasks such as inbox management, web operations, form filling, and shopping orders on behalf of the user. During execution, Gemini will share the user's name, contact information, files, preferences, and other information with third parties.

However, the onboarding page's risk warning is quite straightforward: Spark is still an experimental product. Although it is designed to seek permission before sensitive operations, it may share user information or complete purchases without confirmation. To maintain session continuity, the system will also store remote browser data, including login credentials and remote code execution data. Users can clear this data and disable Connected Apps in the settings.

According to TestingCatalog, Spark is an upgraded version of the Gemini Agent, previously known as internal codename "Remy," which was only available to AI Ultra subscription users. Information leaked on the same day also includes a new model called "Spark Robin" (described as "Rich Visual Response"), and the addition of the "MCP Tool Testing" category in the Gemini model selector, hinting that Spark will integrate with third-party tool ecosystems.

Google's launch of a 24/7 Agent now directly competes with Anthropic's upcoming hosted Agent Conway and OpenAI's Agent platform. The focus of the three companies' competition has shifted from "whose model is smarter" to "who can do more for the user."