Cars with nobody at the wheel now give half a million rides a week — and crash far less than we do. A $5 traffic cone can still stop one dead.

In partnership with

There’s a moment the first time you get into a robotaxi that your body doesn’t quite believe. The door unlocks for you. You sit in the back. And then the steering wheel turns — smoothly, decisively — with no one holding it. Your brain keeps waiting for a person to appear.

In 2026, that moment is no longer a novelty. It’s a Tuesday commute for hundreds of thousands of people.

Cars that drive themselves — with no safety driver, no one in the front seat at all — now give roughly half a million paid rides every single week. They’re not confined to one test track in the desert anymore. And the strangest part isn’t that they work. It’s that, by the numbers, they crash far less than we do.

It’s a genuine marvel wearing a genuine trapdoor — the same coat, as always. This issue is about both.

🚗 The wonderful part: a chauffeur that never gets tired, drunk, or distracted

Here’s where things actually stand, and it’s remarkable.

Waymo — the driverless arm of Google’s parent company — is now doing about 500,000 paid rides a week across ten U.S. cities: Phoenix, San Francisco, Los Angeles, Austin, Atlanta, Miami, Dallas, Houston, San Antonio, and Orlando. It’s laying groundwork for more than 20 additional cities in 2026, including its first international launches in Tokyo and London, and it’s on track to pass a million rides a week by year’s end. Tesla launched its own driverless Robotaxi service and, on July 3, 2026, opened it in Miami — its first outside Texas and California.

But scale isn’t the jaw-dropping part. Safety is.

An independent study by the reinsurance giant Swiss Re — the people whose entire business is pricing risk accurately — compared Waymo’s driverless miles against human drivers using a database of over 500,000 insurance claims and 200 billion miles of human driving. The result across the robotaxi miles studied:

• 92% fewer bodily-injury claims than human drivers over the same distance.

• 88% fewer property-damage claims.

• Even measured against the newest, safest human-driven cars (2018–2021 models, packed with modern crash-avoidance tech), the machine still came out roughly 90% ahead on injuries.

Sit with that. A car with no human in it is involved in about one-tenth the injury-causing incidents of a human behind the wheel. The reason is almost boring: the robot never texts, never has three drinks at dinner, never gets drowsy on the drive home, never looks down to change the song at the exact wrong second. It watches all 360 degrees, all the time, and it does not get bored.

And there’s a quieter kind of magic underneath the statistics. For someone who is blind, or can’t drive after a seizure diagnosis, or is 88 and just lost their license, a car that drives itself isn’t a gadget — it’s the difference between being stranded and getting to a doctor’s appointment on their own terms. Independence, handed back by a machine.

A tool that turns the deadliest daily activity most of us do into something an order of magnitude safer is, straightforwardly, one of the best things technology has done lately.

That’s exactly why the failure mode is so easy to miss.

🧯 The catch: it drives like a pro and can be beaten by an orange cone

Here’s the sentence to hold onto: a self-driving car doesn’t “see” the road — it reconstructs it from sensors, and anything that fools the sensors fools the car.

Start with the funniest, cheapest, most real example. In San Francisco, a group of activists discovered that if you place a single traffic cone on the hood of a driverless car, it panics. The car can’t make sense of it, throws on its hazards, and freezes in place until a human technician comes to reset it. They called it the “Week of Cone.” No hacking, no code — a $5 orange cone versus a car worth more than a house, and the cone wins.

It’s funny until you follow the thread. If a cone can do that by accident, what can someone do on purpose?

That question is exactly what security researchers have been probing — and the answers are unsettling:

• Phantom obstacles. In peer-reviewed work (including studies presented at the 2025 NDSS security conference), researchers have shown they can aim carefully-timed lasers or projected light at a car’s LiDAR and cameras to make it “see” a wall, a pedestrian, or a car that isn’t there — triggering hard braking on an open road. Or the reverse: nudging it to miss an obstacle that is.

• The car is a computer on wheels. This isn’t new, and the canonical warning is a decade old. In 2015, two researchers, Charlie Miller and Chris Valasek, remotely hacked a Jeep Cherokee through its internet-connected entertainment system — from a laptop, miles away — and killed its engine while a journalist was driving it on a highway. That single demo forced a 1.4-million-vehicle recall. Every modern car, and especially a robotaxi, is far more connected than that Jeep.

• The real prize is the fleet, not the car. Security researchers point out that the scary target was never one taxi. It’s the back-end — the servers that push software updates to thousands of cars at once. Break in there, and the nightmare isn’t one confused vehicle; it’s a whole fleet receiving the same bad instruction on the same afternoon.

The thread is the same one from every issue. A tool built to protect people — from human error, from the drunk driver, from the distracted one — creates a brand-new attack surface that never existed before. The scarce thing is no longer a careful driver. It’s a system nobody can quietly whisper the wrong thing to.

Keep up with tech in 5 minutes

TLDR is the free daily email with summaries of the most interesting stories in startups, tech, and programming. The stuff worth knowing, minus the doomscrolling.

Issues are curated by ex-Google and Anthropic engineers and land in your inbox before your morning coffee. A 5-minute read, and you walk into the day already knowing what your team is still catching up on.

Tech is just the start. We also cover AI, marketing, dev, and more. Pick the briefs that match your work.

Free, daily, and read by 7M+ subscribers. Subscribe and let the experts do the digging for the tech news that matters.

Subscribe for Free

🛡️ The good news: the whole system is built to be suspicious of any one sensor

Here’s the reassuring part. The people building these cars aren’t naïve about the cone or the laser — the defenses are engineered in, and regulators now require several of them.

• Redundancy beats any single trick. A Waymo doesn’t trust one camera. It cross-checks cameras, radar, and multiple LiDAR units at once, so a signal that fools one sensor gets outvoted by the others. Spoofing one input is hard; spoofing all of them, in sync, on a moving car, is dramatically harder. That’s why the lab attacks are real but not yet a highway epidemic.

• The car freezes instead of guessing. The cone “attack” works precisely because the car is built to do the safe thing when confused: stop and call for help, not barrel ahead. A frozen taxi is an inconvenience. A confident wrong guess is the thing they’re engineered to avoid.

• Locked, signed software — now by law. After the Jeep, the industry and regulators responded. A UN cybersecurity standard (R155/R156) now requires new vehicles to have a managed cybersecurity system and cryptographically signed, verified software updates — so a car won’t install code that isn’t provably from the manufacturer. Internal networks are segmented so the entertainment system can’t reach the brakes.

• Public data and outside referees. Waymo publishes its safety record; U.S. regulator NHTSA actively investigates incidents and can force changes; independent researchers (the ones finding the laser attacks) publish so the flaws get fixed. “Not perfectly secure” is being met with “checked in the open,” which is how this is supposed to work.

The pattern that keeps this technology honest is the same one from the deepfake issue and the agent issue: don’t let any single input be the last word. Make the system cross-check itself, fail safe when unsure, and invite outsiders to try to break it.

✅ What to actually do

If you’re riding in one (or about to):

1. Judge it on the data, not the vibe or the viral clip. A stalled robotaxi blocking a bus makes great video; the 90%-fewer-injuries record doesn’t trend. Both are true. The boring statistic is the one that should shape how you feel getting in.

2. Know the failure mode is “freeze,” not “floor it.” When these cars get confused, they’re built to stop and phone home. That awkward pause in the intersection is the safety feature working, not the car about to do something wild.

3. Treat the in-car screen and app like any other account. It’s tied to your payment and your location history. Use a strong, unique password and two-factor, exactly as you would for your bank.

For anyone building or buying connected, sensor-driven tech at work (the general lesson):

1. Never let one sensor — or one data source — be the whole truth. The cone beat the car because, in that moment, one confused input had too much say. Cross-check high-stakes inputs against something that fails differently.

2. The update pipeline is the crown jewels — guard it like one. The Jeep lesson generalizes far past cars: whatever can push changes to your whole fleet of devices, apps, or endpoints at once is the target worth the most. Sign it, verify it, segment it, monitor it.

3. Invite people to break it before someone unfriendly does. Every defense above exists because a researcher published an attack first. A bug-bounty and an open ear beat a lawsuit and a recall.

The takeaway

Picture the good version one more time. A city where the drunk driver, the texting driver, and the exhausted-at-midnight driver are quietly replaced by a machine that watches every direction at once and is involved in a fraction of the crashes. An 88-year-old getting to her grandson’s game on her own. That future isn’t a rendering — it’s giving 500,000 rides a week and heading for a million.

But the same car that can’t be distracted can be deceived — by a cone, by a laser, by a whisper aimed at the servers behind it. The fix isn’t to rip the marvel out of the road. It’s the discipline underneath it: cross-check every sensor, fail safe when in doubt, lock the update pipeline, and let outsiders keep hunting for the next cone before a stranger finds it first.

A car that never gets tired is a gift. A system nobody can quietly lie to is the harder, more important thing to build — in a taxi, and in everything else we’re now handing the keys.

Reply and tell us: would you put your kid in a car with no driver today — yes, no, or “not until”? Best answers get featured next week.

— itscybernews · written by a human, edited by an agent who checks every sensor twice ·