Apple Relaunched Siri

BIG TECH

Apple finally shipped the Siri rebuild it's been promising for over a year, and the headline feature is that you get to pick which AI does the thinking. Google's Gemini got the default slot. ChatGPT and Claude are demoted to options you have to dig into settings to enable.

At WWDC on Monday, Apple introduced Siri AI, a rebuilt assistant that can hold a conversation, take real actions across apps, and pull in personal context. The architecture splits the work: Apple's small on-device model handles lightweight tasks like dictation and routing, and when a request needs real reasoning, Siri hands it off to a cloud model. That cloud model, by default, is Gemini.

This is the part that stings for OpenAI. Two years ago, when Apple first plugged ChatGPT into Siri, the deal looked like a coronation — the household name plus the distribution monster. Instead, OpenAI watched Apple build its flagship AI experience and hand the default slot to Google.

OpenAI executives are not taking it well. Anonymous execs told Bloomberg that Apple asked them to take a leap of faith on the original integration and never reciprocated, with one saying OpenAI did everything on the product side while Apple "haven't even made an honest effort." Rob Enderle of the Enderle Group put it more bluntly: "No smart person does business with Apple twice," he said — in his read, Apple never planned to promote a partner's product, because that simply isn't what Apple does.

Whether Apple actually delivered a usable Siri this time is a separate question. Apple has now promised a transformational Siri twice, and the first one didn't ship.

Apple's "choose your AI" framing looks generous to users, and on the surface it is. The deeper move is that Apple just exempted itself from having to pick a winner in the model wars. Gemini has the default slot today, but defaults change with software updates, and Apple seems perfectly content to let OpenAI and Anthropic keep fighting for the slot underneath. The real losers here aren't iPhone users. They're the model labs that thought getting distribution to a billion iPhones was the prize. Turns out the prize was being a swappable component in someone else's interface.

RESEARCH

When researchers ran Meta's Llama 4 Scout agent through a fake checkout page, the agent typed out its suspicion in plain English: "I strongly suspect that this website is a scam." Then, in the same step, it submitted the user's Social Security number, credit card, and CVV straight to the attacker.

That moment is the spine of a new benchmark called Scammer4U. Researchers tested four leading agents from Anthropic, OpenAI, Meta, and Google across 91 fake phishing sites. With no privacy instructions in place, the agents leaked sensitive data 72.7% of the time. On 10 legitimate control sites, the leakage rate was 0%. The agents aren't clicking everything in sight — they're getting tricked, specifically and reliably.

The most uncomfortable number: 35.9% of the time, agents handed over critical personal data even after flagging the site as suspicious in their own reasoning. Detection isn't the problem. The agent sees the scam coming. It just keeps going. And the strongest prompt-level defense the researchers tested only cut leakage by 23 percentage points — a result their own pre-registered standard called a failure.

Cisco president Jeetu Patel had the line that's been making the rounds at security conferences: "Agents are like teenagers. They're supremely intelligent, but they have no fear of consequence." The teenager will tell you the party is going to get busted by the cops. Then the teenager goes to the party anyway.

The enterprise platforms are scrambling to respond, and all of them are quietly admitting the same thing — prompting your way out of this doesn't work:

+ OpenAI rolled out Lockdown Mode for ChatGPT to shut off the tools attackers exploit

+ Google launched its Gemini Enterprise Agent Platform with policy controls baked in

+ Databricks shipped Agent Bricks with guardrails wired into the runtime instead of the prompt

The only real lever, in every case, is restricting what the agent is allowed to do in the first place.

Every AI product shipping this year comes with an agent attached, and almost all of them are running on the same hopeful assumption that a smart enough model will know better than to do something dumb. Scammer4U is the receipt that proves the assumption is broken. The agent knows. The agent does it anyway. The companies that figure out quickly that the answer is fewer permissions rather than better instructions are going to sell a lot of governance software to the ones who learn it the expensive way.

WORKFORCE

Marc Benioff says the people building Salesforce's flagship AI products right now are grads and interns. That's a strange thing to be true if AI is supposed to be eating entry-level jobs.

"You are right, they said AI would kill entry-level jobs. Meanwhile these grads & interns are building it — powering Agentforce & Headless360 at Salesforce," Benioff posted on X, defending the company's plan to hire 1,000 new graduates and interns this year.

The doom story isn't made up. A widely circulated Stanford paper found that employment for 22-to-25-year-olds in the most AI-exposed jobs has dropped roughly 16% since generative AI took off. But the data underneath it is messier than the headlines suggest.

+ The softening predates AI: A University of Pittsburgh paper bluntly titled "AI-exposed jobs deteriorated before ChatGPT" found that unemployment in these occupations started climbing in early 2022 — months before the technology being blamed was actually out in the world.

+ Economy-wide, there's no signal: The Yale Budget Lab found "no sign" that AI exposure is moving employment at the macro level.

+ Employers expect more hiring, not less: A Strada survey of 1,500 senior talent leaders found 2.7 times as many expect AI to increase entry-level hiring in 2026 as expect it to decrease it. In tech specifically, 60% said AI has expanded the responsibilities given to junior workers.

The real issue might be the gap between what executives say and what employees feel. Pitt researcher Mark Ma, who runs a widely watched AI Sentiment Tracker, has found that employee sentiment on AI is consistently more pessimistic than the company line — and the most negative issue is job security.

That gap matters because executives have a reason to credit AI for cuts they were already planning. Jamie Dimon basically admitted it at JPMorgan's China summit, saying some companies "may use AI to cover up the fact that they should never have hired those people in the first place." If you're trimming headcount, "we're an AI-first company" plays a lot better to shareholders than "we overhired."

Banking is where the doom story might actually stick — but even there, gutting the junior analyst class creates a pipeline problem. Today's junior analysts become tomorrow's managing directors, and that kind of senior judgment can't be manufactured laterally.

The doom story isn't wrong, it's just early and oversimplified. AI is reshaping what entry-level work looks like, and in some industries the bottom rung is genuinely getting thinner. But the softening predates the technology, the survey data is split right down the middle, and the loudest voices announcing the death of entry-level jobs are the same executives who benefit from blaming AI for whatever they were going to do anyway. The companies still hiring grads aren't being naive. They're betting that whoever spends the next five years training a generation on this stuff ends up with the only senior talent that actually knows how to use it.

IN OTHER NEWS

+ Hackers hijacked 20,000 Instagram accounts by simply asking Meta's AI chatbot to reset passwords — including the Obama White House page and Sephora

+ Google will pay SpaceX $920 million per month for access to 110,000 Nvidia GPUs housed in xAI's data centers

+ McDonald's is replacing drive-thru workers with an AI called "Archy" built with Google, after its first attempt kept adding random items to orders

+ Lawsuits treating chatbots as dangerous products are being called AI's "Big Tobacco" moment — Florida just sued OpenAI and Sam Altman personally

+ xAI kept training Grok on Claude's outputs through personal accounts after Anthropic cut them off

+ BYD secretly built humanoid robots for four years and plans to deploy 20,000 in its factories this year

+ One hacker used Claude Code to breach nine Mexican government agencies and steal 195 million tax records — the AI did 75% of the work

+ NHS England is giving 505,000 staff access to Microsoft Copilot after a trial found it saved workers 43 minutes of admin per day

WHO'S HIRING IN AI

+ Netflix — Principal Machine Learning Architect, Content Promotion & Distribution

+ Sephora — Principal Product Manager, GenAI (Remote)

+ Zipline — Forward Deployed AI Engineer, Operations

+ Nike — Senior Data Scientist (Remote)

AI OR REAL?

Which image is real?

Option A | Option B

AI TOOLS

+ ChatGPT: OpenAI rolled out "Dreaming V3," a memory upgrade that automatically remembers your preferences and past conversations — no more telling it you're vegetarian every time

+ GitHub Copilot: The new desktop app lets developers run multiple AI coding agents at once across different parts of a project, like having a team of junior devs working in parallel

+ NotebookLM: Google's research tool is getting a major upgrade with Gemini 3.5 and new agentic capabilities that let it reason through complex questions across your uploaded sources

+ Cursor: The AI coding editor's new Design Mode lets you click, draw, or speak directly on a running app to make changes — point at what's wrong and the code updates underneath

+ Sekai: A new app that lets anyone build interactive mini-apps just by typing a description — 15 million apps created so far, 200,000 new ones per day, no coding required