> daily_signal(2026_11_15)
Chicago families rallied to get Google's AI out of their classrooms, in a week that also saw a free AI agent loot a national railway and Mayo put an AI in the exam room.
PickBits Daily Signal · Sunday, November 15, 2026
// tl;dr
- Chicago families rallied outside Google's office to demand a three-year ban on classroom AI. About 30 parents, educators, students and school-board members protested on November 11, part of coordinated actions that also hit Cambridge, Massachusetts. It's the loudest pushback yet to Google's position, which we've tracked since early November, that a school disabling its AI could lose Google Search too, and Chicago Public Schools, the fourth-largest US district, says it's exploring ways to switch the AI off.
- Anthropic says the prompting habits you carried from older models are now slowing Claude Opus 5.5 and quietly costing money. Drop "think hard," start at Medium effort rather than High, and state goal and format in one message. It also tells teams to stop defaulting to cheaper Sonnet, since Opus 5.5 can finish in fewer steps, a cost claim that is Anthropic's own and not independently verified.
- A spy crew used a free, open-source AI agent to run a live espionage campaign against Central Asian governments. CloudSEK's report details how the group Silent Lynx used the Hermes agent to automate parts of an intrusion into Tajikistan's national railway, rode one employee's login into the government document platform, and downloaded thousands of files including defense correspondence.
- Mayo Clinic put an AI clinical assistant at the point of care. A chatbot built on Mayo's own medical knowledge pulls a patient's record and the relevant evidence together during the visit, bundles 22 clinical calculators runnable by voice, and spans more than 35 specialties. There's no independent, published evaluation of its accuracy or effect on outcomes yet.
The Chicago rally is the one I keep coming back to. Early this month we covered Google's line to schools: turn its AI off, and you can lose Google Search with it. I didn't expect the pushback to look like 30 people with signs outside Google's office, but that's where it landed, and Chicago Public Schools, the fourth-largest district in the country, is now openly hunting for a way to switch the thing off.
We've been watching attackers reach for AI agents since the summer, when one hit Thailand's finance sector, so the espionage story isn't a first. What's new is that the agent was free and open-source, the kind of thing anyone can download. Anthropic's prompting update is the small, useful one, though its cost claim is Anthropic's own and worth checking before you believe it. Mayo's exam-room assistant is the one I'm glad to see, even with the proof still missing.
Thirty people stood outside Google's Chicago office this week holding signs that said they didn't want Gemini in their kids' schools.
1. Parents and students rallied outside Google's Chicago office over AI in their schools.
No elected board ever voted on the arrangement they were protesting: keep Google's classroom AI switched on, or lose Google Search.
Google's position toward schools is blunt. A district that turns its AI off can lose Google Search along with it. Nobody got to vote on that. On November 11, about 30 parents, educators, students and Chicago Public Schools board members rallied outside Google's West Loop office, part of coordinated protests that also reached Cambridge, Massachusetts. The rally was organized by Illinois Families for Public Schools and its executive director, Cassie Creswell. Reese Mackey, a 16-year-old junior at Kenwood Academy, spoke against treating AI as the source of answers. The chant was "Hey Google, nice try. We don't want your Gemini."
Look, a search tool that helps a kid check their work is useful. That was never the fight. And Chicago Public Schools, the fourth-largest district in the country, isn't just protesting; it says it's now exploring ways to suppress Google's AI Overview, and may move to another search tool if it can't. Google points to a new "Guided Learning" pilot. The families' list of demands is specific, and any other town could copy it: a three-year moratorium on generative AI in CPS, a named chief privacy officer, screen-time limits across all grades, and a ban on collecting students' biometric data such as face scans and fingerprints.
Why this matters: This one is really about leverage. Google is using one product schools depend on, Search, to keep another switched on in classrooms, and nobody got to vote on that. Thirty people with signs isn't much on its own. The real test is whether a district the size of Chicago can actually refuse, by shutting off the AI Overview or moving to another search provider.
Action this week: Take the Illinois Families demands to your own school board as a template, not a Chicago story: the moratorium, a named privacy officer, the biometric-data ban, the screen-time limits. Ask your district in writing whether its student Google Workspace tier forces AI Overview on, and whether turning it off costs you Search, before someone tells you it does. My read: the board meeting is where the leverage actually sits, because this gets decided in procurement and in the fine print of a Workspace contract, where a parent rarely thinks to look.
2. Anthropic says the prompting habits you carried over from older models are slowing Claude down.
Several of the new tips are the opposite of what power users trained themselves to do.
If your team wrote an internal "how to prompt Claude" guide, there is a good chance it is now wrong, in ways that cost you time and money. On November 14, Anthropic published guidance arguing that habits carried over from older models are holding back Claude Opus 5.5, now the default on its paid plans. Drop the "think hard" and "think carefully" cues: it figures out how hard to think on its own, and in Anthropic's own tests, responses without those phrases started faster with no measured quality loss. Don't set effort to High by default. Start at Medium, where Anthropic says Opus 5.5 matched the previous Opus running on High. Describe the goal, the output format and when to ask questions in a single message instead of spoon-feeding steps. Ask for the finished document, not a draft. Attach charts and tables as files rather than transcribing them.
Whoever owns the AI bill should read the cost claim closely. Anthropic says teams that reflexively route routine work to cheaper Sonnet and reserve Opus for hard problems should reconsider, because Opus 5.5 often finishes in fewer steps, and a heavily configured Sonnet can cost about as much as Opus 5.5 anyway. Those cost numbers are Anthropic's own, though, and the reporting doesn't check them independently. The prompt habits your team wrote down a year ago are probably costing you time or money now.
Why this matters: The habit I'd question first is the reflex to always grab the cheaper model, because that's the one with real money on it. "Use Sonnet for everything routine" felt like a saving; Anthropic is now saying it may not be one. That's a vendor telling you to spend more on its pricier model, so it earns scrutiny. Sure, it's self-serving coming from Anthropic. But how you prompt and which model you pick genuinely cost money now.
Action this week: Delete the "add think hard" and "always set effort to High" rules from your internal prompt guide today, and default new prompts to Medium effort with a single goal-plus-format message. Then get the cost question answered on the record: price a few of your real workloads on Opus 5.5 at Medium against a heavily configured Sonnet, and write down the per-task number yourself rather than taking Anthropic's word for it. When I have watched teams skip that measurement, they end up paying for a "cheaper" model that quietly runs more turns.
Listen to the Daily Signal podcast at pickbits.ai/daily-signal.
3. A spy crew ran an espionage campaign with an off-the-shelf AI agent.
We've watched attackers reach for AI agents since the summer. This time the tool was free, open-source, and aimed at a national railway.
If you run security, the thing to notice is the price: this ran on a free tool. We've seen agent-driven intrusions before; this is the cheap version. On November 11 and 12, Singapore-based CloudSEK published a report, "Off the Rails," on an exposed server, more than a dozen gigabytes of stolen documents, hacking tools and operator records, belonging to an espionage group tracked as Silent Lynx and also called Cavalry Werewolf and YoroTrooper, active since at least 2022. What's new is the tool. The operator used Hermes, a freely available open-source AI agent, configured to act as an offensive-security expert, to automate parts of the operation.
The group held access to Tajikistan's national railway network from mid-August to early October, pulled databases of login credentials, and used a single railway employee's login to enter the national electronic document-management platform that ministries use to exchange official correspondence. From there it built tooling to download thousands of files, including correspondence tied to the Defense Ministry, National Guard and the state security committee. CloudSEK only ties the group to Kazakhstan with medium confidence, though, and found no link to the Kazakh government; and the agent automated parts of the campaign, not all of it. And a real crew still pulled this off with a free agent. That should worry you.
Why this matters: We covered an autonomous agent used against Thailand's finance sector over the summer, and rogue agents spinning up fake identities not long after, so the pattern isn't new. What's changed is how little skill it takes now. The agent here was open-source and free, and it automated the tedious middle of an attack, the reconnaissance, the credential use and the bulk document theft, the work that used to need a human operator's hours.
Action this week: Stand up detection for the shape of this attack now: alert when one account suddenly pulls thousands of files, and when a single credential moves laterally into a document-management system. Scope service and employee accounts tightly, and assume an agent will find an over-permissioned account faster than a human would. Get one answer in writing from whoever runs your document platform: what the alert threshold is for bulk export, and who gets paged when it trips. My own read: the scariest detail is that one stolen login opened a government-wide file store. If you run an internal wiki, picture the same thing happening to you.
4. Mayo Clinic put an AI clinical assistant at the point of care.
It is built to bring a patient's record and the relevant medical evidence together during the visit, instead of sending the clinician off to hunt for both.
You know the moment when a doctor steps out to go look something up? That moment is what this is built for, and the patient sitting there while they search across systems. On November 5, Mayo Clinic Platform launched an AI tool designed to pull a patient's medical information and the relevant clinical evidence together during the interaction itself. It's a chatbot built on Mayo's own medical knowledge base, and it helps a clinician answer clinical questions, explore possible diagnoses and weigh care options, with guidance spanning more than 35 specialties. It bundles 22 clinical calculators, in areas including cardiology, oncology, hepatology and endocrinology, that a clinician can run by typing or by voice, and it integrates into several electronic health records so it can read patient data and help build a care plan.
This is a deployment. Nobody has run the study yet. Mayo says clinicians helped build and test it, and it is available to Mayo staff and to organizations in Mayo Clinic Platform's provider programs, as part of a broader push that includes a June partnership with Microsoft on a healthcare model. What it doesn't come with is an independent, published evaluation of its accuracy or its effect on outcomes. We've been following this one. Back in June we covered MIRA, the system that out-diagnosed a six-doctor panel in a controlled Nature study. MIRA ran in a tidy experiment. Mayo's tool runs in a live exam room, and nobody has measured whether it helps yet.
Why this matters: Mayo frames the goal as giving clinicians their time back, less hunting across resources and more time facing the patient. Take that as their stated aim, not a measured result. What it's good for is narrow and real. An assistant that pulls your chart and the evidence into one place saves time. It's not a second opinion until someone publishes the accuracy data.
Action this week: Watch for the evaluation, not the launch. What I want to see is an independent, published study of whether it changes care decisions or just speeds up lookups. That's the thing that would make it real. Until it lands, use a tool like this to look things up and draft, not to diagnose, and lean on it as a faster way to reach Mayo's evidence and run its calculators. After watching the MIRA result back in June, my instinct is that the in-the-room version is the harder, more useful problem, and I want to see the study before anyone sells it as settled.
healthcaredive.com: Mayo Clinic Platform launches AI tool to streamline clinical decision-making (November 5, 2026)
news-medical.net: AI-enabled Mayo Clinic tool supports individualized patient care decisions (November 2026)
» What to watch this week
- Can a district the size of Chicago actually turn Google's AI Overview off? If it can't, it ends up switching search providers instead, and plenty of other districts are watching to see which way this goes.
- Watch what your own team's AI guidance keeps after this. The real tell is whether anyone actually re-prices a workload on Opus 5.5 at Medium, or just keeps grabbing the cheaper model out of habit.
- Does another security vendor turn up the same thing, an off-the-shelf agent inside a real intrusion? CloudSEK's find is one case. I want to know if it's the first of many, and whether your own document platform could even see one account pulling thousands of files.
- The one I'll be looking for: an independent, published evaluation of Mayo's assistant. Clinicians helping test it is a start; a study is what tells you whether it helps or just feels helpful.
Tomorrow's signal lands here.