AI Is Making Cyberattacks Cheaper, Plus 4 Tech Stories

Monday’s tech news had an accidental theme: AI is becoming less of a weird separate category and more of a layer sitting inside everything else. That can mean useful features, like Google Meet finally giving administrators a real consent control. It can also mean attackers automating more of the boring work involved in breaking into networks.

That second part is where I’m starting today. Taiwan says government agencies were hit by an AI-assisted cyberattack in July, and the important part isn’t that somebody apparently gave a chatbot a black hoodie. It’s that AI can reduce the amount of human labor needed to probe systems, test attack paths and sift through the results. When attacking more targets gets cheaper, being “too small to bother with” becomes an even worse security plan than it already was.

AI-assisted hacking is really an economics story

Taiwan’s Ministry of Digital Affairs says government agencies were targeted in July by an overseas cyberattack that mixed ordinary human hacking with AI agents. Monitoring teams detected abnormal activity, began issuing warnings on July 20 and ultimately contained the incident across the affected agencies.

There’s an important distinction here, because this is exactly the sort of headline that can sprint away from the evidence. Taiwan publicly confirmed an overseas source, but it didn’t name China or another country. Separately, Israeli cybersecurity company Dream said it uncovered an AI-driven operation that stole credentials and personnel information and probed Taiwan’s nuclear-safety agency. Those more specific claims come from Dream’s investigation and related reporting, not from Taiwan’s public confirmation.

So what does “AI-assisted hacking” actually mean? Think less evil robot hacker and more attack team with a pile of extremely fast junior assistants. AI agents can perform reconnaissance, inspect systems, try attack paths, summarize what they find and repeat tedious work in parallel. Humans still choose the target and the objective. The AI makes parts of the process faster and, potentially, much cheaper.

That’s the part that matters outside Taiwan. A small business obviously isn’t a national government, but attack techniques have a way of moving downhill. If more reconnaissance and testing can be automated, attackers can economically poke at more websites, VPNs, Microsoft 365 tenants and employee accounts. You don’t have to become more interesting. You just have to become inexpensive enough to try.

The reassuringly boring answer is that AI hasn’t repealed the fundamentals. Updates, MFA, least privilege, useful logging, good backups and social-engineering defenses still do a remarkable amount of work. AI changes the speed and scale of the problem more than it changes the basic job of defending yourself.

---

Backups & recovery

Backups are comforting right up until you need one. Raymond Tec helps small businesses build practical backup and recovery plans — including the decidedly unglamorous part where we make sure the thing can actually be restored.

---

Apple patches a Screen Sharing flaw worth installing

Speaking of reassuringly boring security work, Apple released macOS Tahoe 26.6.1, Sequoia 15.7.9 and Sonoma 14.8.9 to fix a Screen Sharing authentication flaw. An attacker on the same network could potentially authenticate to Screen Sharing without valid credentials. Apple changed how the authentication state is handled.

The “same network” part matters. This isn’t a report that somebody anywhere on the internet can instantly take over every Mac. But shared networks aren’t exactly rare. Offices, schools, hotels, coworking spaces, apartment buildings and public Wi-Fi all put devices you don’t control in closer proximity to yours than you may realize.

And Screen Sharing is not a feature where I’d get philosophical about whether an authentication flaw is technically serious enough to bother with. The whole point of the service is remote access to a Mac’s screen. If the thing checking credentials can be bypassed, that deserves an update.

As of Monday, there was no public indication that this particular flaw had been actively exploited before the patch shipped. That’s good news, but not a reason to leave the update sitting there. For individuals, install the current update for your supported macOS version. For businesses, make sure managed Macs are actually receiving it instead of assuming somebody will eventually click the button. And if Screen Sharing or Remote Management isn’t needed, there’s no prize for leaving unnecessary remote-access services enabled.

Security updates are boring right up until the moment they aren’t.

---

The boring machines need attention too

Browsers, workstations, remote-access tools, Wi-Fi, and ordinary office hardware rarely get much attention until one of them becomes the problem. Raymond Tec provides onsite IT and field services around Reading, Pennsylvania, along with practical help keeping the technology people use every day working and reasonably secure.

---

Google Meet can finally make consent an actual step

Google Workspace administrators now have a much more sensible option for meetings that are recorded, transcribed or summarized by Gemini: they can require participants to explicitly agree before those features start.

Google first announced the setting in May, paused the rollout in June and resumed it on July 29. The company expected deployment to finish by August 12. The important administrative detail is that the setting is off by default, because of course the privacy-friendly option couldn’t just be the default. An administrator has to enable it at the domain, organizational-unit or group level.

Once it’s turned on, supported participants have to actively consent before Meet can begin “Take notes for me,” recording or transcription. The exact combination depends on the Workspace edition and the features it already includes.

That sounds like one more checkbox buried in an admin console. It’s actually a useful policy tool. Business meetings routinely contain customer information, employee discussions, pricing, project details, health information, legal questions and the occasional sentence nobody expected to become permanent corporate memory. AI note-taking makes capturing all of that wonderfully easy. It also makes it wonderfully easy to forget that everybody else in the room may have a different expectation about what’s being captured.

If your organization uses Meet recording, transcription or Gemini notes, this is worth reviewing. Decide when those tools are appropriate, who can start them and whether explicit consent should be required. The setting doesn’t replace legal, contractual or industry-specific obligations, but it does turn “everybody okay with this?” from an informal courtesy into an actual part of the meeting.

---

Turning on AI is the easy part

Deciding what an AI tool should be allowed to see, who should use it, what work it should perform, and what happens when it gets something wrong is the more interesting problem. Raymond Tec helps businesses connect and automate the tools they actually use without treating every new feature like a button that obviously needs to be switched on.

---

The White House may extend AI security reviews to open models

A federal AI-security framework created in June may soon apply to sufficiently powerful open models as well as the closed systems that were originally at the center of the program.

The underlying policy is already real. A June executive order directed federal agencies to develop classified benchmarks for advanced cyber capabilities in frontier AI models. Once a model crosses the government’s threshold and becomes a “covered frontier model,” developers can voluntarily give the federal government secure access for as long as 30 days before wider release so it can be evaluated for national-security concerns.

The order also says something that’s easy to lose once the argument gets compressed into “government versus open AI”: it explicitly says the framework doesn’t create mandatory licensing, preclearance or a permit requirement for releasing models.

What changed this week is the reported scope. Wired says the White House plans to extend the framework to sufficiently capable open models. That creates a real policy problem because downloadable models don’t behave like services that stay behind one company’s API. Once capable model weights are released and copied around the world, putting the toothpaste back in the tube gets considerably more theoretical.

There are legitimate concerns on both sides. Security officials worry that increasingly capable models can automate cyberattacks and other harmful work. Open-model supporters worry that a framework built around a few giant proprietary labs could become a de facto advantage for exactly those companies, even if the rules remain technically voluntary. Wired also reports concern about how much of the testing system will stay classified, which makes outside evaluation difficult for the obvious reason that outsiders can’t evaluate what they can’t see.

There’s nothing most people or small businesses need to do with this today. The useful questions are narrower: What capability threshold triggers special treatment? What does the review actually test? How voluntary does the process remain in practice? And can enough of it be explained publicly for researchers, smaller companies and everyone else to understand the rules? Those details matter a lot more than whether somebody can fit “innovation” or “safety” onto a podium sign.

---

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.

---

Meanwhile, open-weight AI is becoming a practical business option

That policy debate is becoming more important because open-weight AI is no longer just the cheaper thing sitting several rows behind the frontier models. Reuters reports that cheaper, customizable models are gaining enough traction that major U.S. technology companies are changing course. Meta says it will return to releasing open models, Nvidia is expanding its lineup, and businesses are getting more choices about whether they really need a top-tier frontier model for every prompt, automation or internal task.

“Open-weight” needs translation because it’s frequently used as though it means “open source.” Usually, it doesn’t. An open-weight model makes the trained parameters — the enormous collection of numerical settings learned during training — available for other people to download and run. The company may still keep the training data, training code and other important pieces private. Linux this is not.

Still, downloadable weights can matter a lot. They give developers and businesses more choices about where a model runs, how it’s customized and what happens to the data sent into it. They can also make narrow workloads dramatically cheaper than paying frontier-model prices for every task. If you need a system to classify support tickets, summarize a particular document set or perform another well-defined job, using the smartest model on Earth can be the computational equivalent of taking a tractor-trailer to pick up a gallon of milk.

---

Small custom tools

Custom software doesn’t have to mean a giant multi-year platform. Sometimes a small internal tool that removes one ugly spreadsheet, repeated task, or missing connection is exactly the right amount of software.

---

There’s a catch, naturally. Downloading a model isn’t the same thing as operating one safely. Somebody still has to provide the hardware or cloud service, secure it, update the surrounding software, evaluate its output and understand the license. A cheap model can become expensive remarkably quickly if a business has to build an AI operations department around it.

For a small business, the takeaway isn’t “go self-host an AI model this afternoon.” It’s that the market is becoming more competitive and specialized. Start with the work you need done, the sensitivity of the data involved and the total cost of operating the system. Then choose enough model for the job.

Put all five stories together and the interesting thing isn’t that AI showed up repeatedly. It’s that AI is disappearing into ordinary technology decisions: security, meetings, software costs and government rules. That’s probably where the useful coverage is going to be, too. Less “look, AI!” and more “what did adding AI actually change?”

---

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 / Further Reading

AI-assisted cyberattack in Taiwan

Reuters: Taiwan says it was targeted last month in AI-driven hacking campaign
Taiwan Administration for Cyber Security: AI-assisted ransomware protection guidance

macOS Screen Sharing update

Tom’s Guide: macOS Screen Sharing security update
Golem: Screen Sharing vulnerability and patched macOS versions

Google Meet consent controls

Google Workspace Updates: Require explicit consent for Take Notes with Gemini, recordings, and transcripts in Google Meet
Google Meet Help: participant consent controls

White House AI security framework

White House: Promoting Advanced Artificial Intelligence Innovation and Security
Wired: The White House Is Going to Expand Its AI Policy

Open-weight AI models

Reuters: American AI model makers smell an opportunity