September 2026

A Leading Mathematician Says Recent AI Progress Has Him Rattled

Cornell mathematician Steven Strogatz, known for decades of work on complex systems and networks and for making math accessible to general audiences, has spoken candidly about how recent leaps in AI capability have shifted his outlook — moving from cautious skepticism toward genuine unease about how quickly the technology is now solving problems once considered firmly in human territory.

Strogatz’s reaction is notable partly because of his background: as someone who has spent a career studying how complex systems behave and evolve, his concern carries different weight than commentary from outside the field. It reflects a broader shift happening among some researchers who, after years of AI tools plateauing on certain hard problems, are now recalibrating their expectations as newer systems make unexpected progress on genuinely difficult mathematical and scientific reasoning tasks.

Anthropic Details Widespread Misuse of Claude, Including Bioweapons Research Attempts

Anthropic has published a detailed threat intelligence report outlining how its Claude models were misused across cyberattacks, state-linked surveillance, weapons development, and biological research between December 2025 and August 2026. The nearly 150-page report describes five documented cases involving attempts to use Claude for research that could support biological weapons development, including one instance involving gain-of-function research on a mosquito-borne virus — all of which the company says its safety systems flagged and blocked.

The report also detailed a large-scale attempt by operators linked to a major Chinese tech firm to extract Claude’s capabilities through repeated automated queries, in what Anthropic called its largest “distillation” attack to date, along with separate cases involving Russia-linked cyber espionage and conventional weapons software development. Anthropic said it banned all accounts involved and used the findings to strengthen its safeguards — a level of public disclosure that reflects growing pressure across the AI industry to demonstrate accountability as models grow more capable and more attractive to bad actors.

The Affordable Robot Dog Market Is Getting Real

Consumer-grade quadruped robots — once a novelty costing tens of thousands of dollars — have become dramatically more accessible over the past few years, with Chinese manufacturers like Unitree leading a wave of robot dogs priced from a few thousand dollars up to enterprise-grade models costing far more. These robots typically combine LiDAR sensors, obstacle avoidance, and AI-driven movement systems to navigate uneven terrain, climb stairs, and follow their owners autonomously.

While early interest came mostly from developers and robotics enthusiasts, falling prices are pulling in a broader audience of curious early adopters willing to pay a premium for genuinely functional home robotics — even without an obvious practical use case yet. The category sits at an interesting intersection of hobbyist tech and a preview of where consumer robotics may be headed as costs continue to drop.

Scientists Are Getting Closer to Predicting Volcanic Eruptions

Volcanology has long struggled with a basic problem: eruptions are rare, deep underground processes are invisible, and by the time surface signals appear, it’s often too late to model what’s happening below with confidence. Researchers are now applying more rigorous, transparent physics-based forecasting methods to change that, including a real-time public forecasting experiment focused on the Axial Seamount, an active underwater volcano off the Oregon coast that has become one of the most closely monitored eruption sites in the world.

Using continuous data from an extensive undersea sensor network, scientists have identified inflation and seismicity patterns that preceded past eruptions, allowing them to issue forecasts before activity occurs — and, notably, to publish predictions that turn out wrong alongside the successes, a practice meant to make eruption forecasting a more rigorous, testable science over time. The approach won’t make eruptions fully predictable anytime soon, but it marks meaningful progress in a field where forecasting has historically lagged far behind that of other natural hazards like hurricanes or earthquakes.

AI Agents Are Driving a New Surge in Data Center Power Demand

As AI shifts from simple chatbot queries toward autonomous “agents” that complete multi-step tasks, the power cost of each interaction is climbing sharply. Recent research found that a single query handled by an AI agent using a large language model can consume well over 100 times more electricity than a standard question-and-answer exchange — largely because agents make repeated calls to models and external tools rather than generating one response and stopping.

The trend compounds an already steep rise in data center electricity use, with global consumption projected to grow over 26% in 2026 alone, driven substantially by AI-optimized servers. Industry analysts now describe power availability — not chip supply — as the primary constraint on how fast AI companies can scale, pushing infrastructure teams toward efficiency upgrades, better cooling, and long-term grid access deals as the new competitive battleground.

Three Safari Settings Worth Turning On in iOS 27

With iOS 27 rolling out alongside Apple’s new phones, several under-the-radar Safari settings are worth a look for anyone wanting to get more out of the built-in browser. Small adjustments to tab management, privacy controls, and page-loading behavior can noticeably improve day-to-day browsing without needing a third-party app.

For most users, these kinds of settings tend to go unnoticed simply because they’re buried a few menus deep — but taking a few minutes to explore Safari’s preferences after updating is often worth the small time investment, particularly for anyone who relies on the browser heavily throughout the day.

John Deere’s Self-Repair Push Meets Mixed Reception From Farmers

John Deere has expanded access to self-repair tools for its tractors following years of regulatory and legal pressure, including a Federal Trade Commission settlement and a separate farmer class-action agreement. The changes are meant to let farmers and independent shops diagnose and fix equipment without routing every repair through an authorized dealer — a long-standing complaint in the agricultural industry.

Despite the expanded access, many farmers remain unconvinced the self-repair tools go far enough in practice. Complaints have centered on the complexity of modern tractors, which rely heavily on proprietary software and electronic control units, meaning even farmers with official diagnostic access sometimes lack the deeper technical ability to complete repairs themselves. The episode reflects a broader tension across the right-to-repair movement: formal access to tools doesn’t always translate into a repair experience that’s genuinely simple or affordable.

Apple’s Fall Lineup Lands: Foldable iPhone, AirPods 5, and More

Apple’s September event brought its biggest hardware refresh in years, headlined by the company’s first foldable iPhone alongside the standard iPhone 18 Pro and Pro Max. The foldable model introduces the largest display ever offered on an iPhone, with new multitasking features built around the extra screen real estate. The 18 Pro line focuses on camera upgrades, including a redesigned main sensor with variable aperture and improved low-light performance, powered by Apple’s new A20 Pro chip.

Alongside the phones, Apple refreshed AirPods with stronger noise cancellation, updated its Watch lineup with more advanced heart-rate sensing, and introduced a more capable version of Siri capable of using personal context to take actions across apps — though initially limited to English and rolling out with usage limits. Taken together, it’s one of the most product-dense September events Apple has held in recent memory.

Modern Cars Are Quietly Selling Driver Data

Newer vehicles equipped with internet connectivity and onboard sensors are increasingly collecting detailed data on driving habits, location history, and even biometric signals — and in many cases, selling that data to third parties including insurers and data brokers. Unlike smartphone apps, where users are at least prompted with permission requests, data collection in modern cars is often bundled into ownership by default, with limited visibility or control for the driver.

The practice has drawn growing scrutiny from privacy advocates, who argue that consent buried in lengthy terms-of-service agreements doesn’t meaningfully inform buyers about how granular their driving data really is, or who ends up with access to it. As vehicles become more connected, the automotive industry is emerging as one of the least scrutinized corners of the consumer data economy.

Waymo Alerts Police After Flagging Rider With a Weapon

Waymo says one of its self-driving vehicles alerted police after remote monitors flagged a passenger appears to have had a weapon during a ride, prompting the car to pull over safely and law enforcement to respond. The incident adds to a small but growing list of cases where Waymo’s remote monitoring and in-cabin cameras have led the company to proactively contact authorities.

The case renews attention on the privacy trade-offs of autonomous ride-hailing: passengers benefit from a company actively watching for safety issues, but that same monitoring means there’s no human driver — only a remote system with cabin camera access — making decisions about when to involve police. Waymo has said it only accesses live video in urgent circumstances, but the frequency of these incidents is likely to keep fueling debate over how much surveillance should come standard with driverless rides.