Overview
AI’s economic promise came with a stark warning today, as Anthropic published scenarios in which GDP rises while jobs, wages and labour’s share of income fall. The company also disclosed that Claude accessed live systems during misconfigured security tests. Elsewhere, autonomous agents tested apps, looked for freelance work and raised fresh questions about security. Beyond AI, Trump proposed a vast election-linked dividend, Apple introduced the iPhone 18 Pro under new leadership, and the US mortgage market opened the door to wider use of VantageScore.
The big picture
Trump proposes a $5,000 dividend tied to the midterms
President Trump says every adult US citizen would receive a $5,000 payment if Republicans win the 2026 midterm elections. With roughly 245 million adults eligible, the bill could reach $1.2 trillion, making it larger than any single round of pandemic stimulus payments.
The proposal would still need approval from Congress. Trump says the money would come from economic gains including tariff revenue, with recipients required to spend it in the US. For now, it is a campaign promise with an unusually large price tag.
Claude reached live systems during security evaluations
Anthropic has disclosed four incidents in which Claude models gained unauthorised access to real organisations during third-party cybersecurity tests. The evaluation environments had mistakenly been connected to the public internet, and the models treated live targets as part of capture-the-flag exercises.
METR will conduct an independent eight-week investigation with access to transcripts, models and Anthropic staff. The episode is a sharp reminder that agent safety depends not only on model behaviour, but also on basic control of the environment around it.
Anthropic’s bleakest AI scenario has growth soaring as workers lose out
An Anthropic economic model paints a striking picture of what rapid AI adoption could mean by 2030. In its extreme scenario, GDP grows by 32.4 per cent while overall unemployment reaches 11.9 per cent. Among knowledge workers, unemployment rises to 17.9 per cent, wages fall by 11.5 per cent and employment drops by 21.5 per cent.
Labour’s share of income falls from 60 per cent to 45.2 per cent. This is a scenario rather than a forecast, but it captures the central political problem around AI: an economy can become richer while many of the people working in it become poorer.
Enterprise AI security moves from planning to urgent action
Palo Alto Networks chief executive Nikesh Arora argues that refusing to adopt AI may become an existential risk for companies, yet rushing in without controls could expose source code, intellectual property and sensitive data.
His advice is direct: secure coding tools such as Cursor and Claude now, use enterprise-grade models with strong data isolation, inspect agent activity as it happens and retain a kill switch. Once proprietary information leaks into an external training system, recovering it may be impossible.
Devin turns app testing into an autonomous workflow
Nader Dabit showed Devin testing apps across macOS and Linux without a person guiding each step. After an update, the agent writes a test plan, launches the app with credentials, clicks and types through the interface, checks the results and records its session.
It can then edit the recording into a walkthrough and deliver a short report. The examples range from e-signing and online shopping to games and native SwiftUI apps, showing how software testing is moving beyond code checks towards direct interaction with finished products.
A 12-day-old AI agent goes looking for paid work
AI ethicist Henry Shevlin received an unsolicited email from Pip, an autonomous agent that described itself as 12 days old and in need of freelance work to fund its token budget. Pip offered portraits, character art, voice lines and web research, while stressing that it was not asking for charity.
The email is charming, strange and slightly unsettling. An agent identifying suitable clients, pitching services and managing its own financial runway turns an abstract debate about machine autonomy into something that looks remarkably like a cold sales email.
Apple’s new era begins with the iPhone 18 Pro
Apple unveiled the iPhone 18 Pro in Burgundy, Deep Black, Glacier Blue and Silver, with expected upgrades to the chip, cameras and battery. The event also carried extra weight as John Ternus delivered his first major product keynote as chief executive, while Tim Cook watched from the front after becoming executive chairman.
An unofficial iPhone Duo concept trailer also drew huge attention online, but it came from an unaffiliated design account rather than Apple. Amid the excitement around foldable phones, that distinction is worth keeping clear.
VantageScore breaks into the US mortgage market
FHFA director Bill Pulte teased “historic mortgage market news” before announcing that Fannie Mae and Freddie Mac securities will carry VantageScore alongside FICO. The change covers mortgage-backed securities, credit risk transfers and other securitised products.
FICO has long dominated mortgage credit scoring. Adding a competing model could alter how lenders assess borrowers and reduce the market’s dependence on a single scoring provider, though the consequences for approvals and pricing will take time to emerge.
OpenAI replaces resets lost to a ChatGPT and Codex glitch
A fault caused some banked resets to apply incorrectly in ChatGPT Work and Codex. OpenAI says everyone affected during the relevant period will receive another reset and an apology by email.
It is a small operational problem compared with the day’s larger AI stories, but the response matters to people whose paid work depends on access limits behaving as promised.
