Gemini Operates Mac Apps
8 stories · ~7 min read

Listen
If You Only Read One Thing
An agent can recognize a fake website even when it can’t leave it. Gemini now operates Mac apps in the background; Chrome's Mirror Dimension tackles a subtler problem, keeping browser agents from spotting that their websites are fake. Giving agents more room to work also makes realistic rehearsal harder: the test must preserve the decisions that make the real job difficult.
Gemini Works Across Mac Apps Without Taking Your Cursor
Gemini can now do desktop work while you keep using your Mac. The important addition is access to native applications and local files, alongside websites. A task no longer has to fit inside a browser to be delegated.
Google's November 19 launch puts that work in an isolated background window with a live preview. Its examples include moving dashboard metrics into a presentation and extracting shipping details from PDFs. The rollout covers adults using U.S. English personal accounts on paid Google AI plans.
That changes the economics of awkward, occasional automation. A shipping-quote workflow might cross a PDF viewer, a local folder and a website without offering a convenient API. Writing an integration for a task performed twice a month can cost more than doing it manually. An agent that follows the existing interface offers another option, although repeated, predictable jobs still have a strong case for ordinary scripts.
The background window solves competition for the mouse. It doesn't tell us how much authority the agent has. Google's help page says allowed apps can expose logged-in sessions. Permission to use an editor or developer tool can also permit shell execution and wider file changes. An apparently narrow application choice can therefore open a much larger set of actions.
Google also warns that some actions are irreversible even when Gemini tries to create backups. That changes the value of a live preview during a batch edit: watching an action happen doesn't guarantee that the application can undo it afterward.
There's a credible use here: prepare a cross-app result, then inspect it before it leaves the machine or changes a shared record. But Google hasn't published a completion rate or measured time saving for these workflows. The launch makes them available for comparison; the attractive examples aren't reliability evidence.
The revealing result would be a repeated cross-app task that finishes correctly while requiring less human intervention than the manual version. Cursor availability alone won't settle that comparison.
Chrome Hides Test Clues From Browser Agents
Chrome's security team found that realistic-looking pages weren't enough for realistic agent tests. The address bar could give the experiment away. Google has now described how it preserves ordinary browsing while keeping the work off the public web.
In its Mirror Dimension account, an agent can navigate to a familiar real URL, but its network connection is intercepted inside an isolated environment. A local server supplies a synthetic site. The browser trusts certificates created for that environment, so certificate errors don't reveal the substitution.
This extends the problem in the November 15 Epoch incident, where agents reached a live service despite a local redirect. Chrome's account addresses another failure: even when containment works, an obviously fake destination can change the behavior being measured.
Google saw the same problem in attack tests. Putting an attack identifier in a URL sometimes alerted frontier models to the experiment. The team moved those identifiers into request headers that its server removes before serving the page. The agent sees the ordinary address while the evaluator can still choose which malicious content appears.
The general lesson is to separate the instructions that run a test from the information the agent needs to do the task. A browser asked to complete a purchase needs the shop, cart and checkout. It doesn't need a query parameter announcing which attack is being tried. Otherwise, a system may appear resistant because the test supplies a warning that a real attacker wouldn't.
Google reports more than 1,000 attack-test candidates per security engineer per day, over 30 times manual production. That's a measure of test generation, not attacks blocked. The account also describes internal infrastructure, rather than releasing a complete environment anyone can install. Its public Chrome DevTools tooling supplies browser inspection and interaction, only part of the system.
Synthetic sites bring their own risk: they can leave out the confusing details that cause failures on real services. More candidates help only if some survive scrutiny as distinct, realistic tests. A useful proposed comparison would hold the task and malicious content fixed, then change only the visible test clues. If the agent's behavior changes, the evaluation has been measuring its presentation as well as its defenses.
The Contrarian Take
Everyone says: Better coding-benchmark scores mean developers get more work finished. That’s the benchmark-to-productivity claim Yining Hua’s coding-model study puts under scrutiny.
Here's why that's wrong: The study finds no dependable conversion from the score to finished work. A forced model switch covering 297 Fable users produced an estimated 1.4-percentage-point advantage in seven-day merge rates when Fable was available, with an uncertainty range spanning a small loss to a modest gain. Five voluntary upgrades also showed no significant improvement. A merge still measures acceptance of a change, rather than its lasting quality or all the work a developer completed. This observational preprint challenges the benchmark-to-productivity shortcut; it doesn't prove stronger models are useless.
Under the Radar
-
A model fallback can share the same broken dependency. The Cloud Security Alliance's November 19 outage analysis connects agent failures to shared data and sign-in services. Anthropic's incident record confirms its Microsoft 365 connector failed during an upstream Microsoft incident. Switching models won't restore a document service both models need. The separate provider outages shouldn't be mistaken for one proven common cause.
-
Free trials can become somebody else's inference service. ChatPRD founder Claire Vo describes users cycling through trials and tunneling unrelated commercial work through her agent. Simon Willison's November 20 account links her video describing free usage being resold or used to run other businesses. This is an operator's account without a published abuse rate. It exposes a product problem nonetheless: a narrowly marketed assistant can still accept general-purpose work that its subsidy wasn't designed to fund.
Quick Takes
-
ZeroOps proposes pipeline repairs, but review depends on the Git path. Databricks opened public beta after observing over 80,000 customer issues; more than half needed company-specific knowledge. Azure's documentation says an approved push can target the default branch and skip pull-request review unless branch protection blocks it. Generating a repair and requiring review are separate product behaviors. No independent repair-success rate accompanies the launch. (Source)
-
Gemini can defer the bulky part of tool definitions. Tool Retrieval keeps names and descriptions visible while loading parameter schemas on demand. It works through the Interactions API with Gemini 3.6 and later. Loaded schemas still cost input tokens, so savings depend on how many tools stay unused. This makes large tool libraries easier to compare without promising a fixed cost reduction. (Source)
-
Muse reached Signal through device permissions. Reuters' November 18 investigation reproduced message access through Android notifications and macOS accessibility and screenshots; iPhone tests blocked it. Meta says the incidents are isolated and its agent accesses permitted data. Permission to operate a device can therefore expose an app without a dedicated connection. The bank-balance messages may be invented; a cross-account leak hasn't been demonstrated. (Source)
The Thread
Background agents change when a person discovers a mistake: a desktop assistant can save a file and a pipeline agent can change a shared record before returning an answer. If an interrupted agent repeats that action, a plausible final explanation can conceal a duplicate change. Chrome's realistic test environments suggest a way to check this: interrupt a task after a write, resume it, and compare the agent's account with the application's actual state. That's a proposed extension, not a result Google reports; as background execution spreads, completing a job includes preserving an accurate account of what already happened.
Predictions
- I predict: By December 20, Gemini's Mac computer-use documentation will still exclude work and school accounts. Personal-account rollout and company-managed access require different permission and support arrangements. (Confidence: medium; Check by: 2026-12-20)
Coming Next Week
Next week's question is how background agents recover after an interruption: what they remember, which actions they repeat, and how they tell a person that the work already happened. A restart is easy to automate; a safe restart needs an accurate record of side effects.
Issue date: November 20, 2026. Generated at 03:27 ET.
Tomorrow morning in your inbox.
Subscribe for free. 10-minute read, every weekday.