• Towards AGI
  • Posts
  • Google's AI Can Now Run Your Mac. Who Approves What It Clicks?

Google's AI Can Now Run Your Mac. Who Approves What It Clicks?

your employees' desktops just became agent territory

Today, we're diving into:

  • Gen AI: Google's Gemini can now operate the apps on your Mac, and Apple is reportedly building its own agent while warning users off third-party ones

  • Hot Tea: Citi, PNC and Citizens now say customers carry the losses when a third-party AI agent goes wrong, even if it gets hacked

  • AI News: Meta's Muse agent is reportedly mixing up conversations across accounts and surfacing other people's private data

  • Forward Deployed Watch: EY US is putting Forward Deployed Engineers inside its audit teams

Google's AI Can Now Run Your Mac. Who Approves What It Clicks?

Google has added computer use to the Gemini desktop app for macOS. TechCrunch reported on November 19 that Gemini can now complete tasks by navigating native Mac apps and browsing websites in the background, while you watch a live preview and keep your own screen and cursor free.

Google's suggested uses are ordinary office work: filling out forms, repeating a task across a batch of files, sorting files into folders and building slideshows from files or online dashboards. You can pause, stop and approve actions before they happen. For now it is English only, in the US, for Google AI Plus, Pro and Ultra subscribers.

An Agent That Can Click Anything Needs Rules For Everything.

Find out whether your data and access controls are ready for agents that work inside your employees' apps.

An AI agent operating desktop apps on a laptop while an executive keeps a hand on the pause control

Google is betting that people want an agent to do the clicking for them, and that appetite reaches well beyond the office desktop. That is what OpenMap does, and it is now live: from shopping to planning a group trip, ask in your own words and your agent does the legwork. Try it free at openmap.ai.

The Desktop Is Now Agent Territory

Google is not first here. TechCrunch notes that OpenAI, Anthropic and Meta already offer agentic features in their desktop apps. What matters is that the largest productivity vendors now all want an agent sitting on the same laptop your employees use for email, finance files and customer records.

Apple may be next. According to a Bloomberg report covered by TechCrunch, Apple has posted a job for a senior engineering manager to lead an AI agent that can browse the web, work across apps, run commands and write code.

"Imagine an AI agent that can use a computer the way a capable person does," the Apple job posting reads.

Apple has not confirmed any of this, and Bloomberg says the work is in its early stages. But the same report says the upcoming macOS 27.3 will warn users against giving third-party agents full access to their computers.

Read Those Two Moves Together

One platform owner is making its agent easier to switch on. Another is reportedly making outside agents harder to trust. Your employees will sit in the middle, clicking "allow" on whatever gets the task done fastest.

AI agents queuing at a security gate before they can take control of a company laptop

This Is Not A Productivity Question. It Is A Permissions Question.

The wrong reaction is to treat computer use as a nice feature for power users and let it spread one subscription at a time. An agent that can open any app on a laptop inherits every permission the person logged into that laptop already has.

Picture an analyst who asks an agent to tidy up a folder of quarterly files. The agent does it well. It also opens the shared drive, the CRM tab and the payroll export that happened to be on screen, because nothing told it not to.

The real bottleneck is not whether the agent can do the task. It is whether you have decided which apps, files and accounts an agent may touch, and whether you can see afterwards what it actually did.

The better alternative is to set agent policy at the level of the device and the data, not the individual tool. This fundamentally changes the question from "which agent should we allow?" to "what is any agent allowed to reach?"

What You Should Do Before It Spreads

  • Find out who already uses desktop agents. Check expense reports and app installs for AI subscriptions with computer use.

  • Write a desktop agent policy. Name the apps and data classes no agent may open without approval.

  • Keep approvals switched on. Pause and approve controls only help if people are told to use them for anything that sends, buys or deletes.

  • Log what agents touch. If an agent acted on a file, you should be able to say which one and when.

Before you approve one desktop agent over another, Know Your Inference (KYI) gives you a way to measure it. Test each agent on your own recurring tasks for accuracy, speed and cost, and you will know which one earns a place on your employees' machines, instead of trusting a demo built for someone else's work.

The organizations moving first will not just give employees an agent that clicks for them. They will decide what every agent may reach, and keep a record of what it touched.

Banks Say Your AI Agent's Mistakes Are Your Problem. Who Carries Yours?

Citibank has changed its online user agreement so that customers who hand their banking credentials to third-party AI tools are responsible for whatever happens next, even if the tool is hacked. TechCrunch reported on November 18 that PNC changed its terms the same day and Citizens emailed customers a similar warning.

They are following JPMorgan and Bank of America, which announced similar policies last month. Banks will not be on the hook for reversing transactions that automated assistants make, whether the agent meant to make them or not.

What The Banks Are Saying

  • Outside agents: if a third-party AI tool leaks credentials or is hacked, the customer is responsible for what the attackers do with the account.

  • The bank's own agents: Citi's terms say AI agents Citi itself provides can be used without that extra liability.

  • The reason given: a Citi spokesperson said the bank wants "customers to understand the risks associated with sharing banking credentials."

TechCrunch also points out the commercial angle. Banks see AI agents as a threat to their fees and deposit margins, so the new terms protect the business as well as the customer.

A bank passing the risk of AI agent mistakes back to its customer

Why This Lands On Your Desk

This looks like a consumer banking story. It is really the first clear answer to a question every enterprise will face: when an agent acting for you makes a bad transaction, who pays?

The banks' answer is simple. If you chose the agent, you own the outcome. Expect your suppliers, platforms and insurers to reach for the same wording.

The Wrong Reaction

The easy move is to assume your AI vendor's contract covers you. Most standard terms were written for software that suggests, not software that acts.

The better alternative is to treat agent liability as a contract and controls problem now, before your first agent moves money, signs up for a service or changes a customer record on your behalf.

What You Should Do Now

  • Read your bank and payment terms. Check whether your corporate accounts carry the same AI agent clauses.

  • Map which agents hold credentials. Finance tools, expense bots and procurement assistants often store logins quietly.

  • Prefer read-only access. Give agents the power to analyze before you give them the power to transact.

  • Ask vendors who pays. Put agent errors and agent compromise into your contract reviews and renewals.

The organizations moving first will not just let agents handle money. They will know which agents hold which keys, and who pays when one of them is wrong.

Meta's AI Agent Reportedly Mixed Up Private Chats. Would You Catch Yours?

Meta employees have raised concerns that Muse, the company's AI agent, has been mixing up conversations across different accounts and exposing private data. TechCrunch reported on November 19, citing a Reuters report based on internal company posts.

The examples are uncomfortable. Employees described messages about childcare requests they never made, product orders they never placed and bank balances that were not theirs. Meta did not immediately respond to TechCrunch's request for comment.

What Employees Reported

  • Crossed threads: one employee said Muse added a message listing balances at two brokerage and bank accounts into an unrelated chat about the weather and tides.

  • Unexpected reach: another said they got a reminder about a relative's Signal message, though they had not granted Muse access to other apps.

  • Scale: Muse launched in September and has passed 10 million downloads, according to TechCrunch.

"This seems like an extreme privacy violation," one Meta employee wrote, according to the Reuters report.

These are reported claims from internal posts, and Meta has not responded publicly. Even so, the failure they describe is one every agent builder should recognize.

An AI assistant crossing private messages between different people's phones

This Is Not A Model Problem. It Is A Context Problem.

An agent that reads across apps, chats and accounts builds a working picture of the world from whatever it can reach. If it cannot tell whose record is whose, it will confidently hand the wrong information to the wrong person.

Now picture the enterprise version. A support agent pulls a customer's history, but two customers share a name and the records were never merged properly. The agent answers politely, quickly and with someone else's account details.

The wrong reaction is to blame the model and wait for a better one. The real bottleneck is whether every record your agents read carries a clear owner, a clear identity and a clear rule about who may see it.

What You Should Do This Quarter

  • Test for crossed context. Run agent tests with look-alike customers and shared names, and check nothing leaks between them.

  • Resolve duplicate records first. An agent cannot keep identities apart if your systems cannot.

  • Attach access rules to the data. Permissions should travel with the record, not live only in each app.

  • Treat agent memory as data. Decide what agents may remember, for how long and for whom.

Agents you can trust with private data start with trusted enterprise context. DataManagement.AI helps organizations create that foundation by unifying metadata, lineage, governance, business glossaries and knowledge assets into a single intelligence layer. It resolves duplicate and conflicting records, and ownership and policy travel with every entity, so your agents know whose data they are holding before they act on it.

The organizations moving first will not just give agents more access. They will make sure every agent knows whose data it is holding, and who is allowed to see it.

Forward Deployed Watch: EY Just Put Engineers Inside The Audit

EY US has introduced Assurance Forward Deployed Engineers, a new role inside its audit practice, the firm announced on November 19. The engineers are embedded in engagement teams and combine assurance fundamentals with AI and engineering skills. EY says the move is part of its US firm's $1 billion investment in talent and technology.

The design is worth noticing. The FDEs work alongside the central engineering groups that run EY's enterprise audit platforms, and what they build on engagements can be scaled through its global audit platform, EY Canvas. EY is hiring technologists trained in audit standards and regulatory frameworks, not just engineers.

That last detail is the real story. When the audit profession, whose whole job is evidence, decides AI has to be built by engineers sitting inside the work, it tells you where AI programs actually stall. Not at the demo. At the point where someone has to prove the system works on real data, under real rules, and keeps working.

That is the gap Derisk360 was built for. It embeds trained pods of Forward Deployed Engineers (FDEs), who build governed agentic systems inside your own cloud, alongside Forward Deployed Eval Engineers (FDEEs), who own trust after go-live: eval harnesses, red teaming, guardrails, drift monitoring and audit-ready evidence. Every engineer is trained in the Derisk360 FDE and FDEE Factory. Derisk360 reports that 82% of enterprise AI pilots never reach production without embedded delivery, and that its teams maintain 99.5% agent accuracy in production.

Journey Towards AGI

Research and advisory firm guiding on the journey to Artificial General Intelligence

Know Your Inference

Maximising GenAI impact on performance and Efficiency.

Model Context Protocol

Connect with us, and get end-to-end guidance on AI implementation.

Your opinion matters!

Hope you loved reading our piece of newsletter as much as we had fun writing it.

Share your experience and feedback with us below 'cause we take your critique very critically.

How's your experience?

Login or Subscribe to participate in polls.

Thank you for reading

-Shen & Towards AGI team