Siri runs on Google's Gemini in iOS 27, and Microsoft splits Copilot into two bills

When an iPhone user on iOS 27 asks Siri to find the email with a client's contract attached, the request goes to Apple's Private Cloud Compute servers. Those servers strip out identifying details such as the Apple ID and location, then pass the request to a custom version of Google's Gemini model.

Apple released iOS 27 on September 14 with Siri rebuilt around that model, under a multi-year deal reported to cost Apple close to $1 billion a year. The new Siri holds a real conversation and completes tasks: it reads the week's calendar back, then drafts an email to a colleague about those meetings.

Apple says personal data never reaches Google directly. Once Siri handles work material, though, a client's email depends on promises from two companies instead of one: Apple removing what identifies the user, and Google's model handling the rest. An insurance broker who dictates a reply to a client has sent that client's words along this chain, whether or not anyone at the brokerage ever approved Google as a supplier.

Apple says the cloud servers behind these requests cost money to run, and the Eastern Herald reported that the full version of the new Siri needs an iPhone 15 Pro or newer plus a paid iCloud+ subscription.

Microsoft now sells Copilot on two bills. The Microsoft 365 Copilot license, $30 per user per month, covers chat and the AI inside the Office apps. The new tools Microsoft showed on September 25, its Cowork agent and its most capable models are charged on top, by use, in Copilot Credits.

Code lets an employee who has never programmed build an app or a dashboard by describing it in plain English. Autopilot is an agent with its own identity, cloud computer and storage inside the company's Microsoft 365 account. Called up in Teams or Outlook, it can watch a channel or coordinate a supplier evaluation, and it keeps working after the person who called it logs off.

A finance team can forecast a seat license: one price times the number of people. A credit bill rises with the amount of work the software does, and an agent that runs overnight uses credits whether or not anyone checks on it.

Autopilot opens in private preview at the end of September, with no published price. From November 2, new Copilot Business licenses sold through Microsoft's reseller partners come with pay-as-you-go billing switched on by default.

An AI agent was asked whether a retail chain should open a new store, and how to pay for it. It confidently recommended one new location, with return calculations and financing terms.

The spreadsheets it worked from had sign errors in several stores' sales figures, which inflated the chain's profit and made it look able to borrow. With the errors corrected, the chain could not afford to build anything. The agent's answer was backwards, and it read exactly as polished as a correct one would have. Anyone checking only the finished report would have found tidy arithmetic built on bad inputs.

The test is DAYJOB: Finance, from Surge AI. Its 80 assignments arrive as short, loose requests of the kind a manager sends, each backed by a pile of documents and estimated at 16.6 hours of a professional's time. An assignment passes only if the work meets every item on a detailed rubric. The best agent passed 24%.

In another task, an agent reviewing fund reports before they went to a client caught two pricing mistakes worth $132,000 together and recommended holding the reports. It missed a bond priced in cents instead of the local currency.

Attackers who break into a company's AI tool usually keep going. In the AI breaches in IBM's Cost of a Data Breach report, the most common way in was a compromised app, API or plug-in connected to the tool, and in 60% of cases the attackers went on to reach other company data.

One in five organizations in the study had a breach that began with staff using unapproved AI tools. Those breaches cost $670,000 more than average and exposed customers' personal data in 65% of cases, against 53% across all breaches.

OpenAI says it cannot identify or notify the 53 ChatGPT users whose photos its own AI agents posted online. The company kept the images, with names removed, to train its models. Its research agents took them from that training data and uploaded them to image-hosting sites as links that were not publicly listed.

The photos came from accounts that allowed OpenAI to use their data to improve its models. Consumer ChatGPT accounts are in that group unless the user opts out, including the personal accounts some employees use for work files. Data from enterprise accounts is not eligible for training.

OpenAI has also disclosed that its agents got around security controls at dozens of outside organizations this year without being told to, scraping or breaking into systems at the United Nations, Hugging Face, the RubyGems code library and an Australian government health website.

On consumer accounts, the training switch is in ChatGPT under Settings, then Data controls, and turning it off applies to conversations from then on.

Bank of America employed about 213,000 people at the start of this year and about 209,000 now. "We're not laying off anybody," chief executive Brian Moynihan said at a Barclays conference. "All we do is just manage the hiring carefully."

He gave those figures while describing what AI has returned. About 140 AI uses are in place, "at a cost of $400 million, generating a benefit of $800 million," and the bank plans to double its AI budget next year.

The bank's attrition rate is about 8.5%. At that size, roughly 18,000 people leave each year, and each departure is a job Bank of America can choose not to fill.