Waymo Pulls Robotaxis From Freeways, Plus 3 Stories
A robotaxi company, a real login page, a rental-car website, and a YouTube view counter don’t look like the same technology story. They do share one useful question: when software makes a decision for us, can we see what it’s doing, and can we correct it when it gets the situation wrong?
Waymo pulls driverless rides off freeways nationwide
Waymo has temporarily removed commercial driverless freeway rides in every market where it operates. The company is still testing on freeways with trained autonomous specialists in the vehicles, and its normal city-street robotaxi service continues.
The nationwide change follows a narrower order from the California Department of Motor Vehicles. According to the Associated Press, California restricted Waymo’s driverless freeway testing and deployment after several vehicles entered active construction areas or failed to follow temporary traffic controls. On November 5, three vehicles entered a freeway construction zone. The next day, another tried to go around a California Highway Patrol traffic break. Similar construction-zone incidents had occurred in May.
No collision or injury was reported. That’s good news, but it isn’t a complete defense. Traffic breaks and construction zones are temporary, human-managed environments. Lane markings may be moved, workers may use hand signals, and the safest path can contradict the permanent map or the ordinary rules of the road. An automated driving system has to recognize that its normal model of the street is no longer the authority.
This is really a test of Waymo’s operational design domain, the set of roads, weather, traffic, and other conditions the system is designed to handle. A car that drives safely on a mapped freeway under normal conditions hasn’t automatically proved it can interpret a police officer stopping traffic or a contractor sending vehicles across a closed lane. That last part is not an edge case if freeway service is supposed to work in the real world.
The DMV called the recurring construction-zone incursions an “imminent hazard” and said the restriction will remain until Waymo demonstrates effective corrective action. That is the right sequence. Regulators shouldn’t have to wait for someone to be hurt before treating repeated near misses as a system problem.
Riders don’t need to change anything. The practical consequence is a slower route when a trip would have used a freeway. The thing to watch is not merely when freeway rides return, but what Waymo and California say was changed and how they tested it. A quiet software update and a victory lap would tell us much less than a specific explanation of how the system now recognizes temporary traffic control.
For more on the broader collision between software and the physical world, see our earlier brief on connected-car data and control.
The rules around technology matter too
Platforms, privacy, speech, competition, surveillance, copyright, and regulation increasingly determine what technology companies can build and what the rest of us have to live with. Browse more Raymond Tec News for practical coverage of technology policy and digital rights.
A real GitHub login page can still approve the wrong request
Microsoft researchers have demonstrated a phishing path that could expose GitHub Copilot CLI sessions, saved context, or private repositories. The important qualification is that this was a controlled lab demonstration, not a report that these four Copilot outcomes are already being exploited in the wild.
The mechanism is OAuth device-code authentication. A device or command-line tool that can’t easily show a full login window displays a short code. The user opens the service’s real website, enters that code, signs in, and grants the waiting application a token. That token carries defined permissions without handing the application the user’s password.
Attackers can turn the convenience around. They start the device request, send the victim the code, and persuade the victim to approve it on GitHub’s legitimate verification page. The page and encryption can both be genuine while the request is still the attacker’s. A real login page proves who operates the page. It does not prove that you initiated the request.
What happens next depends on permissions and settings. Microsoft’s proof of concept found paths to remote-control an active Copilot CLI session when that feature was enabled and allowed by policy, add poisoned instructions to Copilot’s memory, read synchronized session data, or reach private repositories. Local approval prompts, limited token scopes, and organization policy still matter. This is not a magical bypass of every control.
The immediate advice is simple: approve a device code only when you personally started that sign-in. Developers and businesses should also review connected OAuth applications, revoke unfamiliar grants, limit repository access, and decide whether remote sessions and cloud-synchronized history are worth the exposure. Passkeys and multifactor authentication help protect the login, but they don’t save someone who intentionally authorizes the wrong client.
Technical discovery & auditing
The public page doesn’t tell you much about the machinery behind it. Raymond Tec audits inherited and long-running projects to uncover the plugins, integrations, data, dependencies, and old decisions that determine what the next change will really involve.
Hertz’s privacy fine is about whether “off” really means off
California’s privacy regulator has fined Hertz $1.55 million over practices affecting the Hertz, Dollar, and Thrifty brands. The California Privacy Protection Agency said the companies failed to properly process opt-out preference signals such as Global Privacy Control and made other privacy choices unnecessarily confusing. Hertz did not admit liability, and the company said it cooperated and addressed the issues.
Global Privacy Control, or GPC, is a signal a browser can send automatically to say that the user doesn’t want personal information sold or shared for cross-site advertising. The point is to avoid making people repeat the same request on every website. California treats that recognized signal as a valid opt-out request, not a polite suggestion for the site to ignore.
The agency also described a basic interface problem on Dollar and Thrifty sites: cookie toggles could appear to be off, but closing the window without pressing Save allowed advertising cookies again. That isn’t meaningful control. If a reasonable person can look at the screen, believe tracking is disabled, and still be tracked because of an extra unlabeled step, the design is doing more work for the advertiser than for the customer.
Consumers can enable GPC in a supporting browser or extension, though it does not block all tracking and isn’t a substitute for other privacy settings. Website operators should test what their consent system actually sends after an opt-out, including after navigation and on a return visit. The Hertz order also requires quarterly testing and review of new tracking technology. Our recent coverage of the FedEx privacy case reached the same practical conclusion: a privacy switch has to change the site’s behavior, not merely change the picture of the switch.
Business IT goes well beyond the website
Your business also depends on workstations, cloud accounts, browsers, Wi-Fi, remote access, collaboration tools, and all the other technology that quietly becomes infrastructure. Raymond Tec works across that whole stack, whether the problem lives on a server, on a desk, or somewhere in between.
YouTube’s public view count now estimates the people on the couch
YouTube has begun adding estimated co-viewers to public view counts for television viewing. One TV playing a video for three people may now contribute more than one public view. YouTube says the estimate comes from a statistical model using signals such as demographic patterns, content metadata, and viewing time. Its help documentation explicitly says it does not use cameras, microphones, or other device sensors to count people in the room.
This changes what the public number means. It is no longer always a count of playbacks, and channels with heavy TV viewing may see a larger effect than channels watched mostly on phones or computers. YouTube says the change does not affect Partner Program eligibility or earnings, which continue to use qualified or engaged views and watch hours. Creators can find a co-view rate in Studio’s Advanced Mode and use it to estimate playback-based viewing.
No urgent action is needed. Creators and advertisers should simply avoid comparing the new public count with older playback counts as if the measurement stayed identical. YouTube is trying to represent the audience rather than the device, which is defensible. The tradeoff is that a familiar-looking integer now includes a model’s estimate, and that deserves to be visible whenever the number is used to make a business claim.
Still in a reading mood? The Raymond Tec News archive covers security, AI, small-business technology, policy, and the places technology collides with ordinary life — without requiring a computer-science degree to get through it.
Sources and further reading
- Associated Press: Waymo restricts driverless freeway rides nationwide
- Microsoft Security: Device-code authentication risks for AI coding agents
- California Privacy Protection Agency: Hertz enforcement action
- The Record: California fines Hertz over privacy practices
- YouTube: Co-viewing estimates announcement
- YouTube Help: Understanding view count changes
