Crystal ball: Just a couple of months after Cloudflare revealed that bots have officially surpassed human traffic online, the company has given its view on what things will look like in five years. It’s not good news for us fleshy meatsacks: non-human traffic could be up to 1,000 times higher than human traffic.

CDN and cybersecurity giant Cloudflare held its Q2 earnings call on Thursday, during which chief financial officer Thomas Seifert gave a concerning prediction about what the web will look like in the future.

“If the current trends continue, we think in five years, non-human traffic will be as much as 1,000 times as much as human traffic,” he said.

“In other words, humans will be a rounding error on the internet, not because human traffic goes down, but that’s just how fast we’re seeing non-human traffic grow.”

Seifert did add the important caveat that his predictions have been wrong in the past. Cloudflare CEO and co-founder Matthew Prince will likely be able to relate to this. He had predicted that bot traffic would surpass humans at the end of 2027, but it happened earlier this year.

Also read: The Zero Click Internet

It goes without saying that the sudden acceleration in non-human traffic is being driven by AI – agentic AI, specifically. Agentic systems behave very differently from traditional web crawlers and fraud bots. Their activity can closely resemble ordinary browsing, but it happens at machine speed and can be repeated on an enormous scale.

Even a straightforward prompt may trigger thousands of requests. Prince illustrated the difference using online shopping: someone searching for a camera might check five retailers, while an AI agent could examine 5,000 sites on their behalf. Multiply that behavior across millions of users asking assistants to compare products and airfares, research topics, collect information, or condense articles, and you can see why Seifert believes humanity’s online presence will become a rounding error.

Seifert said that if traffic does grow the way he expects, “we’ve got to get a lot more efficient,” which is where Cloudflare’s services come in, of course. He also mentioned the security implications of a web made up pretty much entirely of bots.

The idea that Dead Internet Theory – the concept that the internet is mostly run by bots and AI – will be demonstrably true in a few years (or now, arguably) is a concerning one. For a start, it could see the end of independent sites that have relied on the web’s business model since the 1990s.

There’s also the fact that an internet full of bots, AI slop, and privacy risks is making more people pull away from the online world. In our poll, 97% of people said the internet today was a hellscape compared to its early days, and it looks like things are only going to get worse.

Alibaba released Qwen 3.8-Max this week and marketed the preview as second only to Claude Fable 5 (their launch-day table was more equivocal: the model leads on one of 12 coding-agent rows). But an independent harness came close to the opposite conclusion: a benchmark run, apparently using the Preview version, put Qwen 3.8-Max’s best effort setting mid-pack, and its default setting last.

Both results are real and defensible. The gap between them is about token and time budgets, and that matters because those figures aren’t usually headline numbers. Alibaba’s footnotes give its coding numbers a five-hour timeout, and up to 12 hours per run on PaperBench. The independent harness, VulcanBench, allowed between 45 and 60 minutes of wall clock time. A time budget between five and 16 times larger on Alibaba’s side explains the huge difference in results.

It’s time to do two things to start accounting for these differences when choosing models. First, the metric to use is cost per successful task: total spend, including everything you spent on attempts that failed, divided by the tasks that actually passed your acceptance check. Second, you need to make time or token budgets an explicit part of your acceptance criteria, not a hidden detail.

Price per token has stopped predicting the bill

The comparison everyone published in Qwen 3.8-Max’s first week was a price comparison, because that was the only data available. It is not a cheap model. DeepSeek-V4-Flash-0731, which entered public API beta on July 31, lists at 14 cents per million input tokens and 28 cents output. Qwen 3.8-Max lists at $2 and $6. Kimi K3 sits at $3 and $15.

Those prices tell you less than they used to, for a reason specific to reasoning models like Qwen: getting to a result costs thinking tokens. A model that spends most of its token allowance on reasoning can reach a token cap before it writes the answer, giving you an empty result indistinguishable from a total failure at the cost of a full run.

Artificial Analysis has the cleanest published measurement of how this can affect real agent spend: running its Intelligence Index on DeepSeek-V4-Flash at maximum effort took 210 million output tokens against a class median of 100 million. Absolute cost stayed low anyway, because the tokens were so cheap. But verbosity costs time, not just money, and depending on your use case that can sink you.

What you need is a number that counts everything you spent, including the attempts that came back empty, against the tasks that actually got done in the time and token budget you specified. This is what a cost-per-success metric helps you see.

Your failure rate is partly a configuration setting

A run that produces a wrong answer and a run that runs out of budget are different events with different fixes. Almost no harness distinguishes them, and almost no leaderboard reports the split. I hit this building an agent benchmark of my own: the harness logged a failure and nothing about why, and I had to add the distinction myself. When you do separate them, budget exhaustion turns out to dominate.

Long-Horizon-Terminal-Bench, published in July, ran 17 frontier models across 46 tasks through a shared harness with one 90-minute attempt each. Timeouts accounted for 79% of unresolved runs, against 19% for agents that stopped on their own and 3% for harness errors. The authors are careful about what that does and does not mean: the timed-out runs were not close to finishing, with mean reward between 0.10 and 0.35, so you cannot assume more time would have resulted in success. But the lesson is: benchmarks are implicitly measuring time efficiency, whether or not they shout about that.

The clearest published example of the mechanism comes from VulcanBench, the same open-source harness behind the Qwen chart. In a report dated July 26, Claude Opus 5’s lowest-effort setting was its best, solving 20 of 23 tasks against 18 at high effort. The extra reasoning wasn’t useless: high effort returned the fewest wrong answers of any setting, one against three. It ran out of clock instead, and a timeout scores zero. Two of its three regressions were cutoffs on tasks that low effort solves, and given unlimited time on both it only ties its cheapest setting, at 3.1 times the cost.

That has a direct consequence for anyone building a routing ladder. The standard design escalates to more reasoning when a cheap attempt fails, on the assumption that the next rung is better and merely costs more. For a meaningful share of model and task combinations that assumption is wrong, and you pay the higher rung’s price to escalate into a timeout or hitting a cap.

Who is already measuring this

Several groups have landed on cost per successful task independently in the last few months, which is the strongest signal it’s becoming standard.

VulcanBench reports dollars per solved task as a headline column and has since its earliest reports. Long-Horizon-Terminal-Bench publishes per-task cost next to accuracy, and its most instructive row is GPT-5.4 at roughly $26 per task with a much lower pass rate than Grok 4.5 at about $11. TestEvo-Bench runs agents under a cost cap, and Claude Code’s test-generation score falls from 71% to 44% at the tighter cap.

Vendors are already on board with the idea of measuring per successful task. HubSpot moved its Breeze Customer Agent in April to 50 cents per resolved conversation, down from $1 per handled conversation. Zendesk bills per automated resolution. Fin charges 99 cents per outcome and bills only on end-to-end resolution.

What to change this week

• Emit a failure reason on every agent run as a required field, with budget exhaustion, verifier failure and harness error as distinct values rather than one failure flag. Until you can separate a timeout from a wrong answer, your pass rate is measuring two things at once and you cannot tell which one to fix.

• Compute cost per successful task per effort level, not just per model. Total spend including failed attempts, divided by tasks that passed your acceptance check. The ranking will not match the rate card, and the cheapest setting may well win.

• Cap on tokens rather than wall clock unless latency is genuinely in your service level objective. A wall-clock cap scores your provider’s serving speed as model quality.

• Check the default effort setting on everything you have deployed. Qwen 3.8-Max runs at its highest reasoning setting when the effort field is unset, and its highest setting was its worst performer in independent testing. A team that never touches that parameter is running the configuration that costs the most per solved task.

If your vintage gaming system eats through batteries in a hurry, this is for you.

The right rechargeable batteries

Panasonic / IKEA

USB-C battery mods

Handheld Legend

RingConn has always been a solid challenger to Oura in the smart ring space, and after spending two weeks testing it, I can safely say the new RingConn Gen 3 continues that trend. You’ll get most of the same health, sleep and recovery tracking features as rivals in our best smart rings guide here, but without the monthly subscription that puts some people off buying a smart ring altogether.

The Gen 3 tracks heart rate, heart rate variability (HRV), blood oxygen levels, skin temperature, respiratory rate, stress, activity and sleep. New additions include a haptic motor for some alerts and reminders, improved battery life and new vascular health trend insights, which are designed to give you a picture of your cardiovascular wellbeing over time.

For smartphones selling at over $600, Apple owns 65% of that market and grew 9% year over year in the first half of 2026. However, Chinese brands are making a dent in local market share.

Apple ships 23% of all smartphones sold globally and captures 49% of all profit. It’s a behemoth on its own, but when accounting for smartphones in the $600 plus range, it’s the clear leader in market share.

According to a report from Counterpoint Research, Apple captured 65% of the global premium smartphone market. It grew 9% year over year during the first half of 2026, though its market share is still down when compared to the 74% held in 2022.

The decline since 2022 can be attributed to the premium Chinese smartphone market growing in popularity. When a Chinese consumer is presented with an option between an iPhone and a smartphone from Huawei, Xiaomi, or Oppo, they choose the local brand.

That said, Apple still holds most of the market globally at 65% and Samsung at 19%. The other OEMs make up the remaining 16%.

Apple holds onto 65% of premium market share even as the rest of the market increases prices. Image source: Counterpoint

The premium smartphone market grew 5% overall year over year for the first half of 2026 and makes up a record 29% of the global market. The global RAM shortage driving up prices has pushed more smartphones into the premium $600 plus segment and price hikes on lower-end models are driving customers to more premium models.

The report suggests these trends are why smartphone manufacturers are leaning further into financing options. Samsung recently launched its Galaxy Card while Apple has introduced Apple Upgrade.

Trade-in prices have also gone up, as was seen with Apple earlier Thursday. Everything is meant to entice customers holding onto their more expensive devices for longer to upgrade sooner in spite of price increases.

It’s going to be an interesting fall as everything seems to be creating quite the hurdle for Apple to overcome. iPhone 18 Pro sales could be limited by the RAM shortage and Apple reports that it’s already selling as many iPhone 17 models as it can make.

Demand won’t be the issue, but supply held back by global supply chain instability. Consumers may be feeling the impact of economic uncertainty, but that may not matter if financing options continue to be plentiful.

CYBER-CRIME

Cancer diagnostics breach spills personal and health info as extortion crew says healthcare giant ‘should’ve paid the ransom’

Hackers have dumped data stolen from Abbott’s cancer diagnostics business after the healthcare giant apparently declined to pay up, with the leak containing 10.9 million unique email addresses alongside personal and health information.

Have I Been Pwned added the Exact Sciences breach to its database on Friday after data stolen by the ShinyHunters extortion crew was published online. The leak includes names, email and physical addresses, phone numbers, dates of birth, genders, and personal health information belonging to customers, patients, and healthcare providers.

Abbott, which acquired cancer diagnostics outfit Exact Sciences earlier this year, first disclosed the break-in on July 16. In an update on August 5, it acknowledged that some of the files accessed contained personal information and/or personal health information, but said it was still analyzing the data and had yet to determine who needed to be notified.

The company also revealed that the intrusion began with a vishing attack, and stressed that this was “not an encryption malware event.” It said only a limited number of internal systems within its cancer diagnostics business were affected, with no disruption to products, manufacturing, laboratory operations, or patient services.

ShinyHunters has a rather different way of describing what happened. On its dark web leak site, seen by The Register, the extortion crew told Abbott it “should’ve paid the ransom” and claimed the company had failed to reach an agreement despite being given multiple chances to do so.

The crooks claim that they made off with more than 30 million rows of customer information, including more than one million Social Security numbers and 7.5 million dates of birth. More concerning are claims that the haul includes 22 million-plus rows of client notes containing confidential doctor-patient conversations and health information, as well as more than 20 million medical-order records containing patient IDs, prescription types, order dates, and refill information.

ShinyHunters also claims to have siphoned more than 425 million rows of assorted data from Databricks, 130,000 files from SharePoint, and 89,000 contracts from Coupa, though those figures have not been independently verified.

Abbott has not said how the vishing attack led to the theft of the data, how long the intruders had access to its systems, or whether it received an extortion demand. Its latest update says the investigation remains ongoing and that affected individuals will be notified where required.

For those caught up in the breach, however, that review may come a little late. Whatever number Abbott eventually puts on the incident, a sizeable chunk of the stolen data is already in the wild. ®

For more than a decade, the cybersecurity industry has been assigning names to different hacking groups. Some of them, like Fancy Bear, have crossed over into the mainstream because of their prominent hacks and memorable names. Others are only known within the cybersecurity industry. 

Oftentimes, even industry insiders can’t keep track. In part, that’s because every company names hacking groups differently. That’s why there are resources like this one, which attempt to be a one-stop shop where cybersecurity professionals, government officials, policymakers, journalists, and the wider public can make sense of who is who. 

Last month, Google became the latest company to revamp its naming system for hacking groups.

Gone are the days APT1, APT41 or APT whatever number, which was the system adopted by Mandiant, once an independent security firm that’s now part of Google. Mandiant was the first to adopt a naming scheme.

From now on, Google’s system is relatively simple: A hacking group will have a first name that is memorable and random, and a second word whose initial indicates the country of origin: Castle for China, Ion for Iran, Neptune for North Korea, and Relic for Russia.

According to Shane Huntley, the chief technology officer of Google Threat Intelligence Group, the company’s in-house hacker hunting team, the revamp was necessary to bring clarity to security researchers both inside the company and externally. 

In the early 2010s, when companies started publishing reports on cyberattacks and naming the hackers behind them, Huntley told TechCrunch that, “we were not expecting to have as many threat groups as we do today.”

It had become hard to keep track of everyone. Google now tracks more than 5,000 “activity clusters” in several countries, according to John Hultquist, chief analyst at Google Threat Intelligence Group. Huntley said that there are very few developed nations that don’t have their own cyber capabilities and hacking groups. 

But what is the point of naming hacking groups? It’s not just an academic exercise, Huntley explained. The goal is to have a baseline understanding of who is attacking who, and how they are attacking them. That way organizations can recognize threats more quickly, prepare against them, ideally stop them, or at least investigate incidents more promptly. 

All that, he said, it’s possible only if you name the hackers and track them consistently. 

“If you actually get hacked by them or you’re dealing with some incident, knowing how that actor behaves, what they do, what they’ve done in the past, all of these details become critically important to help the response and also work out your coverage against these threats as well,” said Huntley.

Knowing how the North Korean government hackers known as the Lazarus Group behaves, what their goals usually are, and who they work for, gives defenders a starting point in dealing with these hackers. 

Tracking state-sponsored hackers, while challenging, is easier than tracking cybercriminal groups and hackers-for-hire, Huntley explained. The government hackers tend to have more consistent targets and activities, while cybercriminal groups have members that come and go, sometimes splinter, and otherwise are more amorphous. Hacker-for-hire groups and spyware makers tend to have a lot of customers in different parts of the world, making them slightly harder to track. 

A common criticism whenever a new naming system gets announced is: Why don’t all companies and organizations just use the same codenames? While that seems like an easy question to answer, the reality is that every company has a slightly different view of every group, based on their own sets of data and telemetry. Huntely said this is an inescapable reality that can’t be avoided just by sharing more information among companies and groups of researchers. 

“No one has perfect visibility,” he said. “We are building our model and our best understanding, but we will never know everything about what’s going on.”

By unifying the naming scheme of Google’s old Threat Analysis Group, which Huntely headed, and Mandiant, at least now there’s one fewer scheme to remember. For everything else, refer to this gargantuan list.