ESET tracks rise in malicious AI skills and adaptable malware
ESET spent the first half of 2026 looking at AI skills, the plug-ins that AI agents use to do actual work. They found something worth your attention: Out of around 900K skills that they analyzed, tens of thousands were suspicious and thousands were just flat-out malicious.
It's the browser store extension problem all over again. Small executables published by whoever, installed by users who want a job done, running with whatever permissions the agent has. We have been down this rodeo before. It ran for a decade in the Chrome Web Store and it ended with locked extension installs and an allowlist.
The difference now is that nobody has built the allowlist yet. If your district has AI agents in your environment, and you probably do whether you approved them or not, there is no equivalent to the extension policy you have in Google Admin. Go ahead, ask your vendors what skill vetting looks like, the answer will tell you alot.
ESET also identified PromptSpy, which they said is the first known Android malware to use gen AI inside its execution flow. It uses Gemini to read UI elements to figure out what to tap, rather than using hardcqoded coordinates and screen names. Malware that adapts to whatever changes the OEM makes to their buttons.
It is rare, and ESET says LLM guardrails are probably holding adoption back. Fine. Guardrails erode. The interesting part is not that PromptSpy exists but that the technique works at all, which means signature-based detection on mobile just got a longer to-do list.
The vendor report ends where these always end: trust the vendor. Read it anyway, then go check what your agents are allowed to install.
Featured articles
Closing the Visibility Gap in K–12 IT With Observability
A teacher can't load the assessment platform. Is it the AP in that room, the district firewall, the SSO provider, the vendor's own cloud, or the kid's Chromebook? EdTech's pitch is observability: one view across on-prem gear, SaaS, and endpoints instead of six dashboards that each swear everything is green. IDC's Matthew Leger points out the gaps show up exactly where traffic hands off between systems you own and systems you don't, which is the part of the diagram nobody monitors. The honest version of this is that you already know where the blind spots are. You just can't prove it yet.
Security Awareness Training Reduces Phishing Success
IBM's figure is that human error plays a role in roughly 95% of breaches, which is the kind of number that gets used to sell you a product. But EdTech's piece on awareness training is worth reading anyway, mostly for Randy Rose at CIS reframing the problem. Teachers aren't careless. They're processing a firehose of information and expected to decide fast. That's not a character flaw, it's a working condition.
Which is why simulated phishing campaigns that name and shame click-throughs tend to backfire. You're not fixing the environment, you're punishing people for living in it.
Opinion: Why Claude for Teachers Might Be a Student Privacy Trap for Educators
Anthropic's Claude for Teachers is free to any U.S. K12 educator who applies with a school email address and asserts the account is for educational purposes. That is the whole gate. No district agreement, no DPA, no signature from anyone who has read your state's student data privacy statute.
Andrew Rice and Mark Schneider make the obvious point in The 74: a tool that onboards teachers individually is a tool that routes around you entirely. The Learning Commons integration and OpenSciEd alignment are genuinely good. That is exactly why it will spread through your staff before you hear about it.
Your school email domain just became a procurement signature.
Adults have struggled to set rules for AI in school. These teens figured it out
Ninety-eight high schoolers, one per state and change, spent a late-July weekend in a replica Senate chamber drafting AI legislation for K12 schools. They got further in three days than most district AI committees have gotten in three years.
Worth reading not because teenagers have solved anything, but because they asked the questions cleanly: AI on homework, AI on tests, whose job it is to teach any of this. Those are the questions that eventually land on your desk as a filtering rule, a Chrome extension whitelist, or an angry email about a flagged essay.
The students had the advantage of not needing to please anyone's school board.
Tech tip of the week
Set up network segmentation for IoT devices to isolate potential malware spread
Your smart TVs, HVAC controllers, security cameras, and that one lonely digital sign in the cafeteria are all sitting on a network that can reach your file servers. Most of them run firmware nobody has patched since installation, and several will never be patched again because the vendor stopped caring in 2019.
Put them somewhere else. Build a dedicated IoT VLAN, tag the switchports those devices live on, and write the ACL so the VLAN can reach DNS, DHCP, NTP, and whatever cloud endpoint the device actually needs. Deny everything else, especially east-west traffic between IoT devices themselves. Cameras have no reason to talk to thermostats.
Start with one building and one device class. Cameras are usually easiest because their traffic patterns are boring and predictable. Log the denies for a week before you enforce, since something always needs a port you did not expect.
The point is not to make IoT secure. It never will be. The point is that when one of those cameras gets popped, the blast radius is one VLAN.
Pass This Along
If a colleague or coworker in K12 IT would find this useful, forward it. They can subscribe free and get their own copy every Monday.