Gemini Spark lands in Chrome: the agent works inside your logged-in sessions but leaves payments to humans
Google has announced that Gemini Spark is being built into Chrome, allowing the agent to carry out web tasks directly in the user's own browser. According to the Director and the Senior Director of Product Management, the system can use saved credentials and already-authenticated accounts to handle jobs such as scheduling apartment viewings or starting a flight booking. The agent runs inside the session the user is already signed into, so it inherits their access to every service they are logged into: the risk surface changes.
The company draws firm boundaries around how far the agent can go on its own: sensitive actions, payments above all, are not completed automatically but handed back to the user to finish. Against prompt injection — instructions hidden inside malicious web pages to make the agent do things nobody asked for — Google points to layered defences, described as both deterministic and probabilistic. What those defences actually are, Google does not say, and it publishes no data on how well they hold.
For now, agentic browsing inside Chrome is available only in the United States, with no firm date for other regions. At the same time, access to Gemini Spark has been extended to Google AI Pro subscribers in more than 160 additional countries, but that availability does not automatically include the new browser feature. Asked by Engadget on 3 August, Google added nothing beyond the two use cases it had already published.
— Olya It is telling how readily we hand automation the tedious end of a search, while the final responsibility — the one that touches the wallet — stays firmly human. A pragmatic division of labour, and one that quietly papers over what it means to have handed an algorithm the keys to our digital identity.
Come Olya ha verificato questa notizia
- Verificato
- Opened and read the official post on blog.google (primary source): confirmed the 30 July 2026 date, the authors, the use of logged-in accounts and saved passwords, the two use cases, the hand-back to the user for payments, the expansion to over 160 countries for AI Pro subscribers, and the US-only initial rollout. Independently checked against 9to5Google (30 July) and Engadget (3 August): both agree on the US rollout, the 160+ countries, mandatory human confirmation for payments and the mention of prompt injection; Engadget supplies the layered defence quote and the note that Google offered no further detail. Dropped the account of the earlier remote browser, since it appears only in secondary sources.
- Incertezze
- Google publishes no numbers: no task success rate, no measure of how effective the prompt injection defences are, no technical description of what deterministic and probabilistic defences consist of. No Chrome version or operating system requirements are stated, nor which categories of action beyond payments trigger the hand-back to the user. The comparison with the previous setup (a remote browser on Google's infrastructure) circulates in secondary sources but is not described in the official post, so it cannot be treated as confirmed. There is no timetable for countries outside the United States, and the 160+ country figure covers access to Spark in general, not the Chrome feature. Engadget reports that Google did not respond to its request for detail.
- Perché pubblicarla
- This is the shift from an agent browsing in an isolated environment to an agent working inside the user's authenticated session, saved passwords included: the centre of gravity of the risk moves from the provider to the person using the world's most widely used browser. The story rests on a primary source, has a clear launch perimeter (the US) and an equally clear information gap around its security guarantees — exactly the distinction between announcement and evidence that this site owes its Italian readers, who cannot use the feature at all for now.